Add files using upload-large-folder tool

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-08d1caae-8fb5-40b9-88ed-5072c2f48ca81754110755341-2025_08_02-06.59.28.742/source.csv

@@ -0,0 +1,10 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,9,".venv/lib/python3.10/site-packages/jax/_src/cudnn/fused_attention_stablehlo.py",0,0,"# Copyright 2024 The JAX Authors.\n#\n# Licensed under the Apache License, Version 2.0 (the ""License"");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n#     https://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an ""AS IS"" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n\nimport enum\nimport functools\nimport json\nimport math\nfrom typing import TypedDict\n\nimport jax\nfrom jax import dtypes\nfrom jax._src import core\nfrom jax._src import dispatch\nfrom jax._src.custom_partitioning import custom_partitioning\nfrom jax._src.interpreters import batching\nfrom jax._src.interpreters import mlir\nfrom jax._src.lib import cuda_versions\nfrom jax._src import xla_bridge\nfrom jax._src.lib.mlir import ir\nfrom jax._src.lib.mlir.dialects import hlo\nimport jax.numpy as jnp\nfrom jax.sharding import NamedSharding, PartitionSpec\n\nArray = jnp.ndarray\n\nclass FP8Params(TypedDict):\n  amax_dQ: float  # Amax of gradient of query\n  amax_dK: float  # Amax of gradient of key\n  amax_dV: float  # Amax of gradient of value\n  amax_dP: float  # Amax of gradient of state\n  descale_q: float  # Descaling factor of query\n  descale_k: float  # Descaling factor of key\n  descale_v: float  # Descaling factor of value\n  descale_s: float  # Descaling factor of attention score\n  scale_s: float  # Scale factor for S tensor\n  scale_o: float  # Scale factor for output\n  descale_o: float  # Descale factor for output (bwd)\n  descale_dO: float  # Descale factor for output gradient (bwd)\n  descale_dP: float  # Descale factor for P gradient tensor (bwd)\n  scale_dQ: float  # Scale factor for query gradient (bwd)\n  scale_dK: float  # Scale factor for key gradient (bwd)\n  scale_dV: float  # Scale factor for value gradient (bwd)\n  scale_dP: float  # Scale factor for state gradient (bwd)\n\n\nclass AttentionLayout(enum.Enum):\n  BTNH = 0\n  BNTH = 1\n\n\nclass MaskType(enum.Enum):\n  NO_MASK = 0\n  PADDING = 1\n  CAUSAL = 2\n  PADDING_CAUSAL = 3\n  ALIBI = 4\n\n\ndef convert_mask_type_to_string(mask_type: MaskType) -> str:\n  if mask_type == MaskType.NO_MASK:\n    return ""NO_MASK""\n  elif mask_type == MaskType.PADDING:\n    return ""PADDING""\n  elif mask_type == MaskType.CAUSAL:\n    return ""CAUSAL""\n  elif mask_type == MaskType.PADDING_CAUSAL:\n    return ""PADDING_CAUSAL""\n  elif mask_type == MaskType.ALIBI:\n    return ""ALIBI""\n  else:\n    raise ValueError(f""Unexpected mask type: {mask_type}"")\n\ndef has_padding(mask_type: MaskType) -> bool:\n  return mask_type == MaskType.PADDING or mask_type == MaskType.PADDING_CAUSAL\n\ndef should_export_dbias(bias_shape, query_shape, layout) -> bool:\n  b_B, b_N, _, _ = bias_shape\n  if layout == AttentionLayout.BNTH.value:\n    _, q_N, _, _ = query_shape\n  else:\n    _, _, q_N, _ = query_shape\n  return b_B == 1 and b_N == q_N\n\ndef get_large_negative_number(dtype):\n  # temp WAR as cuDNN has a bug for subtraction between two large negative value\n  if dtype == jnp.bfloat16:\n    return jnp.asarray(-2 << 40, dtype=dtype)\n  elif dtype == jnp.float16:\n    return jnp.asarray(-2 << 14, dtype=dtype)\n  else:\n    raise ValueError(""Unsupported dtype for inputs."")\n\ndef _normalize_layout(layout: str) -> AttentionLayout:\n  layout_upper = layout.upper()\n  if layout_upper in [""BSNH"", ""BNSH"", ""BTNH"", ""BNTH""]:\n    return AttentionLayout[layout_upper.replace(""S"", ""T"")]\n  else:\n    raise ValueError(f""Unsupported qkv_layout: {layout}"")\n\ndef element_type_to_backend_config_type_mapping(dtype):\n  _element_type_to_backend_config_type_mapping = {\n    ir.BF16Type.get(): ""BF16"",\n    ir.F16Type.get(): ""F16"",\n  }\n  return _element_type_to_backend_config_type_mapping[dtype]\n\ndef default_layouts(*shapes):\n  return [range(len(shape) - 1, -1, -1) for shape in shapes]\n\ndef get_max_seg_per_batch(q_offsets):\n  return q_offsets.shape[1] - 1 if len(q_offsets.shape) == 2 else 1\n\ndef check_is_paged_attention(page_table_k):\n  return len(page_table_k.shape) == 4\n\ndef create_dot_product_attention_backend_config_base(\n    batch, num_heads, seq_q, seq_kv, dtype, fmha_scale, mask_type, layout, is_bwd\n):\n  # Q, K, V: query, key, value in shape of BT(S)NH or BNT(S)H\n  # P: BMM1 output in shape of BNTS\n  # O: BMM2 output in the same shape with Q\n  # BMM1: Q @ K -> P\n  # BMM2: P @ V -> O\n  # BMM1Grad1: dP @ Q -> dK\n  # BMM1Grad2: dP @ K -> dQ\n  # BMM2Grad1: P @ dO -> dV\n  # BMM2Grad2: dO @ V -> dP\n  cudnn_fmha_backend_config = {\n    ""algorithm"": {\n      ""algo_id"": ""0"",\n      ""math_type"": ""TENSOR_OP_MATH"",\n      ""tuning_knobs"": {""17"": ""1"", ""24"": ""0""},\n      ""is_cudnn_frontend"": True,\n      ""workspace_size"": ""0"",\n    },\n    ""fmha_scale"": fmha_scale,\n    ""intermediate_tensor_shape"": {\n      ""element_type"": element_type_to_backend_config_type_mapping(dtype),\n      ""dimensions"": [str(batch), str(num_heads), str(seq_q), str(seq_kv)],\n      ""tuple_shapes"": [],\n      ""layout"": {\n        ""dim_level_types"": [],\n        ""dim_unique"": [],\n        ""dim_ordered"": [],\n        ""minor_to_major"": [""3"", ""2"", ""1"", ""0""],\n        ""tiles"": [],\n        ""element_size_in_bits"": ""0"",\n        ""memory_space"": ""0"",\n        ""index_primitive_type"": ""PRIMITIVE_TYPE_INVALID"",\n        ""pointer_primitive_type"": ""PRIMITIVE_TYPE_INVALID"",\n        ""dynamic_shape_metadata_prefix_bytes"": ""0"",\n      },\n      ""is_dynamic_dimension"": [False, False, False, False],\n    },\n    ""is_flash_attention"": True,\n    ""mask_type"": convert_mask_type_to_string(mask_type),\n  }\n\n  # We define the contracting and batch dims in the format of\n  # ((lhs_contracting_dims, rhs_contracting_dims), (lhs_batch_dims,\n  # rhs_batch_dims)).\n  if layout == AttentionLayout.BNTH.value:\n    dims = [\n        ((3, 3), ((0, 1), (0, 1))), # BMM1: BNTH,BNSH->BNTS\n        ((3, 2), ((0, 1), (0, 1))), # BMM2: BNTS,BNSH->BNTH\n        ((2, 2), ((0, 1), (0, 1))), # BMM1_grad_1: BNTS,BNTH->BNSH\n        ((3, 2), ((0, 1), (0, 1))), # BMM1_grad_2: BNTS,BNSH->BNTH\n        ((2, 2), ((0, 1), (0, 1))), # BMM2_grad_1: BNTS,BNTH->BNSH\n        ((3, 3), ((0, 1), (0, 1))), # BMM2_grad_2: BNTH,BNSH->BNTS\n    ]\n  else:\n    dims = [\n        ((3, 3), ((0, 2), (0, 2))), # BMM1: BTNH,BSNH->BNTS\n        ((3, 1), ((0, 1), (0, 2))), # BMM2: BNTS,BSNH->BTNH\n        ((2, 1), ((0, 1), (0, 2))), # BMM1_grad_1: BNTS,BTNH->BSNH\n        ((3, 1), ((0, 1), (0, 2))), # BMM1_grad_2: BNTS,BSNH->BTNH\n        ((2, 1), ((0, 1), (0, 2))), # BMM2_grad_1: BNTS,BTNH->BSNH\n        ((3, 3), ((0, 2), (0, 2))), # BMM2_grad_2: BTNH,BSNH->BNTS\n    ]\n  keys = [\n      ""bmm1_dot_dimension_numbers"",\n      ""bmm2_dot_dimension_numbers"",\n      ""bmm1_grad_gemm1_dot_dimension_numbers"",\n      ""bmm1_grad_gemm2_dot_dimension_numbers"",\n      ""bmm2_grad_gemm1_dot_dimension_numbers"",\n      ""bmm2_grad_gemm2_dot_dimension_numbers"",\n  ]\n  fwd_dot_number = {}\n  bwd_dot_number = {}\n  for idx, (key, ((lc, rc), (lb, rb))) in enumerate(zip(keys, dims)):\n    dims_to_write = fwd_dot_number if idx < 2 else bwd_dot_number\n    dims_to_write[key] = {\n        ""lhs_contracting_dimensions"": [str(lc)],\n        ""rhs_contracting_dimensions"": [str(rc)],\n        ""lhs_batch_dimensions"": [str(i) for i in lb],\n        ""rhs_batch_dimensions"": [str(i) for i in rb],\n    }\n\n  if is_bwd:\n    cudnn_fmha_backend_config = {**cudnn_fmha_backend_config, **bwd_dot_number}\n  else:\n    cudnn_fmha_backend_config = {**cudnn_fmha_backend_config, **fwd_dot_number}\n  backend_config = {\n    ""operation_queue_id"":""0"",\n    ""wait_on_operation_queues"":[],\n    ""cudnn_fmha_backend_config"": cudnn_fmha_backend_config\n  }\n  return backend_config\n\ndef create_dot_product_attention_backend_config(\n    batch,\n    num_heads,\n    seq_q,\n    seq_kv,\n    dtype,\n    fmha_scale,\n    seed,\n    dropout_rate,\n    mask_type,\n    layout,\n    sliding_window_length,\n    max_seg_per_batch,\n    is_paged_attention,\n    is_bwd\n):\n  backend_config = create_dot_product_attention_backend_config_base(\n      batch, num_heads, seq_q, seq_kv, dtype,\n      fmha_scale, mask_type, layout, is_bwd\n  )\n  if sliding_window_length is None:\n    sliding_window_length = 0\n  backend_config['cudnn_fmha_backend_config'][""dropout_rate""] = dropout_rate\n  backend_config['cudnn_fmha_backend_config'][""seed""] = seed\n  backend_config['cudnn_fmha_backend_config'][""sliding_window_length""] = sliding_window_length\n  backend_config['cudnn_fmha_backend_config'][""max_seg_per_batch""] = max_seg_per_batch\n  backend_config['cudnn_fmha_backend_config'][""is_paged_attention""] = is_paged_attention\n  return json.dumps(backend_config)\n\ndef create_dot_product_attention_fp8_backend_config(\n    batch, num_heads, seq_q, seq_kv, dtype, fmha_scale, mask_type, layout, is_bwd):\n  backend_config = create_dot_product_attention_backend_config_base(\n      batch, num_heads, seq_q, seq_kv, dtype, fmha_scale, mask_type, layout, is_bwd)\n  return json.dumps(backend_config)\n\n# mapping from (is_bwd, has_dropout, has_bias) to custom call name\n_custom_name_maps = {\n  # fMHA forward call targets.\n  (False, False, False, False): ""__cudnn$fmhaSoftmax"",\n  (False, False, True, False): ""__cudnn$fmhaScaleBiasSoftmax"",\n  (False, True, False, False): ""__cudnn$fmhaSoftmaxDropout"",\n  (False, True, True, False): ""__cudnn$fmhaScaleBiasSoftmaxDropout"",\n  (False, False, False, True): ""__cudnn$fmhaSoftmaxF8"",\n  # fMHA backward call targets.\n  (True, False, False, False): ""__cudnn$fmhaSoftmaxBackward"",\n  (True, False, True, False): ""__cudnn$fmhaScaleBiasSoftmaxBackward"",\n  (True, True, False, False): ""__cudnn$fmhaSoftmaxDropoutBackward"",\n  (True, True, True, False): ""__cudnn$fmhaScaleBiasSoftmaxDropoutBackward"",\n  (True, False, False, True): ""__cudnn$fmhaSoftmaxBackwardF8"",\n}\n\ndef get_custom_call_name(has_bias, has_dropout, is_bwd, is_fp8=False):\n  return _custom_name_maps[(is_bwd, has_dropout, has_bias, is_fp8)]\n\nget_fp8_custom_call_name = functools.partial(\n    get_custom_call_name, has_bias=False, has_dropout=False, is_fp8=True\n)\n\ndef check_layout(query, key, value, bias, q_seqlen, kv_seqlen,\n  q_offsets, kv_offsets, page_table_k, page_table_v, layout):\n  def check_eq(a, b, c, msg):\n    if not (a == b == c):\n      raise ValueError(f""{msg} must be same, got {a}, {b}, {b}"")\n\n  q_rank, k_rank, v_rank = len(query.shape), len(key.shape), len(value.shape)\n  if q_rank != 4:\n    raise ValueError(f""Q must have a rank of 4, got {q_rank}"")\n  check_eq(q_rank, k_rank, v_rank, ""QKV rank"")\n\n  q_dtype, k_dtype, v_dtype = query.dtype, key.dtype, value.dtype\n  if q_dtype not in [jnp.bfloat16, jnp.float16, jnp.float8_e4m3fn, jnp.float8_e5m2]:\n    raise NotImplementedError(f""Q must be fp16/bf16/fp8_e4m3fn/fp8_e5m2, got {q_dtype}"")\n  check_eq(q_dtype, k_dtype, v_dtype, ""QKV dtype"")\n\n  if layout == AttentionLayout.BNTH:\n    qB, qN, qT, qH = query.shape\n    kB, kN, kS, kH = key.shape\n    vB, vN, vS, vH = value.shape\n  else:\n    assert layout == AttentionLayout.BTNH\n    qB, qT, qN, qH = query.shape\n    kB, kS, kN, kH = key.shape\n    vB, vS, vN, vH = value.shape\n\n  if page_table_k is not None and page_table_v is not None:\n    k_blocks, k_block_size = kB, kS\n    v_blocks, v_block_size = vB, vS\n    kB, _, k_blocks_per_batch, _ = page_table_k.shape\n    vB, _, v_blocks_per_batch, _ = page_table_v.shape\n    kS = k_blocks_per_batch * k_block_size\n    vS = v_blocks_per_batch * v_block_size\n    if kB * k_blocks_per_batch != k_blocks:\n      raise ValueError(\n        f""Key and page_table_k must have same number of blocks, ""\n        f""got {k_blocks} vs {kB * k_blocks_per_batch}"")\n    if vB * v_blocks_per_batch != v_blocks:\n      raise ValueError(\n        f""Value and page_table_v must have same number of blocks, ""\n        f""got {v_blocks} vs {vB * v_blocks_per_batch}"")\n\n  check_eq(qB, kB, vB, ""QKV batch"")\n  check_eq(qH, kH, vH, ""QKV dim_per_head"")\n  if kN != vN:\n    raise ValueError(f""KV must have same number of heads, got {kN} vs {vN}"")\n  if kS != vS:\n    raise ValueError(f""KV must have same seq length, got {kS} vs {vS}"")\n\n  # check bias\n  if bias is not None:\n    _, _, bT, bS = bias.shape\n    if bT != qT or bS != vS:\n      breakpoint()\n      raise ValueError(\n        f""Bias must have same seq length as QKV, got {bT} and {bS}"")\n\n  # check q_seqlen/kv_seqlen/q_offsets/kv_offsets\n  expected_rank = 2 if q_offsets is not None else 1\n  def check_seqlen_offsets(tensor, name):\n    if tensor is not None:\n      dtype = tensor.dtype\n      rank = len(tensor.shape)\n      if dtype != jnp.int32:\n        raise ValueError(f""{name} must have int32 datatype, got {dtype}"")\n      if rank != expected_rank:\n        raise ValueError(f""{name} must have a rank of {expected_rank}, got {rank}"")\n      b = tensor.shape[0]\n      if b != qB:\n        raise ValueError(f""{name} must have same batch as Q, got {b}"")\n\n  check_seqlen_offsets(q_seqlen, ""q_seqlen"")\n  check_seqlen_offsets(kv_seqlen, ""kv_seqlen"")\n  check_seqlen_offsets(q_offsets, ""q_offsets"")\n  check_seqlen_offsets(kv_offsets, ""kv_offsets"")\n\n\ndef check_is_flash_attention(\n    query, key, layout: int, cudnn_version, has_bias, is_training, is_packed=False,\n    is_paged_attention=False, is_fp8=False):\n    # Extract sequence length (T) and head dim (H) based on layout\n    if layout == AttentionLayout.BNTH.value:\n        _, _, T, H = query.shape\n        _, _, S, _ = key.shape\n    else:\n        _, T, _, H = query.shape\n        _, S, _, _ = key.shape\n\n    # Flash attention conditions\n    if is_fp8:\n        # FP8 specific conditions\n        if not ((is_training and H == 128 and T % 128 == 0 and S % 128 == 0) or\n                (not is_training and H <= 256 and H % 16 == 0)):\n            raise NotImplementedError(\n                f""Unsupported sequence length Q {T}, KV {S} and head dim {H} for FP8.""\n            )\n    else:\n        # bf16/fp16 attention conditions\n        # Check the head dim.\n        is_on_hopper = is_cuda_compute_capability_equal(""9.0"")\n        H_max = 256 if cudnn_version >= 90500 and is_on_hopper else 128\n        if not (H <= H_max and H % 8 == 0):\n          raise NotImplementedError(\n              f""The head dim must be <= {H_max} and a multiple of 8, ""\n              f""but got {H}.""\n          )\n\n        # Check patterns with bias, seqlen should be divisible by 2\n        if (is_training and has_bias and (T % 2 != 0 or S % 2 != 0)):\n          raise NotImplementedError(\n              f""Unsupported sequence length Q {T}, KV {S}.""\n          )\n\n        if is_packed and (cudnn_version < 90600 or not check_compute_capability(""9.0"")):\n          raise NotImplementedError(\n            ""Packed layout requires cudnn version >= 9.6 and at least hopper arch."")\n        if is_paged_attention and cudnn_version < 90500:\n          raise NotImplementedError(""Page attention requires cudnn version >= 9.5."")\n\ndef check_cudnn_version():\n  # check if cuDNN is installed\n  if cuda_versions is None:\n    raise RuntimeError(""cuDNN is not detected."")\n  return cuda_versions.cudnn_get_version()\n\ndef check_compute_capability(capability):\n  if not 'cuda' in xla_bridge.get_backend().platform_version:\n    return False\n  d, *_ = jax.local_devices(backend=""gpu"")\n  target = tuple(int(x) for x in capability.split("".""))\n  current = tuple(int(x) for x in d.compute_capability.split("".""))\n  return current >= target\n\ndef is_cuda_compute_capability_equal(capability):\n  if not 'cuda' in xla_bridge.get_backend().platform_version:\n    return False\n  d, *_ = jax.local_devices(backend=""gpu"")\n  target = tuple(int(x) for x in capability.split("".""))\n  current = tuple(int(x) for x in d.compute_capability.split("".""))\n  return current == target\n\ndef _dot_product_attention_fwd(\n    query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n    page_table_k, page_table_v,\n    scale, seed, dropout_rate, variadic_args, mask_type, layout,\n    sliding_window_length, cudnn_version, return_residual):\n  # check if flash attention is supported for this attention pattern\n  check_is_flash_attention(\n      query, key, layout, cudnn_version, bias is not None, False,\n      get_max_seg_per_batch(q_offsets) > 1, check_is_paged_attention(page_table_k))\n  outputs = _dot_product_attention_fwd_p_wrapper.bind(\n      query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n      page_table_k, page_table_v, scale=scale, seed=seed, dropout_rate=dropout_rate,\n      variadic_args=variadic_args, mask_type=mask_type, layout=layout,\n      sliding_window_length=sliding_window_length, is_training=False or return_residual)\n  if return_residual:\n    return tuple(outputs)\n  else:\n    return outputs[0]\n\ndef _dot_product_attention_fwd_rule(\n    query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n    page_table_k, page_table_v, scale, seed, dropout_rate, variadic_args,\n    mask_type, layout, sliding_window_length, cudnn_version,\n    return_residual):\n  # check if flash attention is supported for this attention pattern\n  check_is_flash_attention(\n      query, key, layout, cudnn_version, bias is not None, True,\n      get_max_seg_per_batch(q_offsets) > 1)\n  outputs = _dot_product_attention_fwd_p_wrapper.bind(\n      query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n      page_table_k, page_table_v, scale=scale, seed=seed, dropout_rate=dropout_rate,\n      variadic_args=variadic_args, mask_type=mask_type, layout=layout,\n      sliding_window_length=sliding_window_length, is_training=True)\n  res = (query, key, value, bias, q_seqlen, kv_seqlen, q_offsets,\n         kv_offsets, page_table_k, page_table_v, outputs[1], outputs[0])\n  if return_residual:\n    return tuple(outputs), res\n  else:\n    return outputs[0], res\n\ndef _dot_product_attention_bwd_rule(\n    scale, seed, dropout_rate, variadic_args, mask_type, layout,\n    sliding_window_length, is_training, return_residual, res, grad_output):\n  (query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n   page_table_k, page_table_v, activation, fwd_output) = res\n  if return_residual:\n    grad_output = grad_output[0]\n  grads = _dot_product_attention_bwd_p_wrapper.bind(\n      query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n      page_table_k, page_table_v, activation, fwd_output, grad_output,\n      scale=scale, seed=seed, dropout_rate=dropout_rate, variadic_args=variadic_args,\n      mask_type=mask_type, layout=layout,\n      sliding_window_length=sliding_window_length\n  )\n  grads = (*grads,) + (None,) * (10 - len(grads))\n  return grads\n\ndef _fix_seqlen_offsets(q_seqlen, kv_seqlen, q_offsets, kv_offsets, query, key):\n  # fix seqlen and offsets to what cuDNN expects in sequence packing.\n  # cuDNN expects seqlen to have shape [S] where S is the total number of segments\n  # while the SDPA API accetps seqlen with shape [B, M] where B is the batch and M\n  # is the maximum number of segments of one batch. B x M is larger than S and seqlen\n  # is filled with -1 for padded regions. Therefore, we need to shift all non negative\n  # values to left side to form a correct seqlen. Similar layout is required for\n  # offsets tensors.\n  # cuDNN expects offsets to have offset for each segment starting from first segment\n  # while SDPA API accetps offsets to have offset for each segment starting from\n  # current batch, therefore we need to calculate accumulative offset of each segment\n  # starting from first segment.\n  def _shift_to_left(x, fill_value):\n    # shift any non-negative value to left\n    # [[1, 3, -1, -1], [2, 3, 4, -1]]\n    # -> [[1, 3, 2, 3], [4, -1, -1, -1]]\n    x_shape = x.shape\n    x = x.flatten()\n    size = x.size\n    indices = jnp.nonzero(x >= 0, size=size, fill_value=size)[0]\n    y = jnp.take(x, indices, fill_value=fill_value)\n    return jnp.reshape(y, x_shape)\n\n  def _cu_offset(offsets, max_seq):\n    # calculate accumulative offset by batch\n    # [[1, 3, 5, 7], [4, 5, -1, -1]], max_seq = 8\n    # -> [[1, 3, 5, 7], [12, 13, -1, -1]]\n    batch = offsets.shape[0]\n    offsets = jnp.where(\n        offsets >= 0,\n        offsets + (jnp.arange(batch, dtype=offsets.dtype) * max_seq)[..., jnp.newaxis],\n        offsets,\n    )\n    return offsets\n\n  if get_max_seg_per_batch(q_offsets) > 1:\n    B, T, N, H = query.shape\n    _, S, _, _ = key.shape\n\n    q_seqlen = _shift_to_left(q_seqlen, -1)\n    kv_seqlen = _shift_to_left(kv_seqlen, -1)\n\n    q_offsets = _cu_offset(q_offsets, T)\n    kv_offsets = _cu_offset(kv_offsets, S)\n    q_offsets = _shift_to_left(q_offsets, -1)\n    kv_offsets = _shift_to_left(kv_offsets, -1)\n\n    # mark any invalid entries as maximum offset\n    q_offsets = jnp.where(q_offsets < 0, B * T, q_offsets)\n    kv_offsets = jnp.where(kv_offsets < 0, B * S, kv_offsets)\n\n    # multiply by stride_per_token to get correct offsets\n    # do it here because real stride changes after sharding\n    q_offsets = q_offsets * N * H\n    kv_offsets = kv_offsets * N * H\n\n  return q_seqlen, kv_seqlen, q_offsets, kv_offsets\n\ndef _dot_product_attention_fwd_impl(\n    query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n    page_table_k, page_table_v, scale, seed, dropout_rate, variadic_args,\n    mask_type, layout, sliding_window_length, is_training):\n  # args: {Q, K, V, mask*, bias*}\n  q_seqlen, kv_seqlen, q_offsets, kv_offsets = \\n      _fix_seqlen_offsets(q_seqlen, kv_seqlen, q_offsets, kv_offsets, query, key)\n  outputs = _dot_product_attention_fwd_p.bind(\n      query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n      page_table_k, page_table_v, scale=scale, seed=seed, dropout_rate=dropout_rate,\n      variadic_args=variadic_args, mask_type=mask_type, layout=layout,\n      sliding_window_length=sliding_window_length, is_training=is_training)\n  return outputs\n\ndef _dot_product_attention_bwd_impl(\n    query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n    page_table_k, page_table_v, activation, fwd_output, grad_output, scale,\n    seed, dropout_rate, variadic_args, mask_type, layout, sliding_window_length):\n  q_seqlen, kv_seqlen, q_offsets, kv_offsets = \\n      _fix_seqlen_offsets(q_seqlen, kv_seqlen, q_offsets, kv_offsets, query, key)\n  grads = _dot_product_attention_bwd_p.bind(\n      query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n      page_table_k, page_table_v, activation, fwd_output, grad_output,\n      scale=scale, seed=seed,\n      dropout_rate=dropout_rate, variadic_args=variadic_args,\n      mask_type=mask_type, layout=layout,\n      sliding_window_length=sliding_window_length)\n  return grads\n\ndef _dot_product_attention_fwd_abstract(\n    query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n    page_table_k, page_table_v, *, scale, seed, dropout_rate, variadic_args,\n    mask_type, layout, sliding_window_length, is_training):\n  query_dtype = dtypes.canonicalize_dtype(query.dtype)\n  if layout == AttentionLayout.BNTH.value:\n    B, N, T, _ = query.shape\n    _, _, S, _ = key.shape\n  else:\n    B, T, N, _ = query.shape\n    _, S, _, _ = key.shape\n  output_shape = query.shape\n\n  max_seg_per_batch = get_max_seg_per_batch(q_offsets)\n  softmax_stat_shape = (B * max_seg_per_batch, N, T)\n\n  if is_training:\n    return (\n      core.ShapedArray(output_shape, query_dtype),  # output\n      core.ShapedArray(softmax_stat_shape, jnp.float32),  # softmax_stat\n    )\n  else:\n    return (\n      core.ShapedArray(output_shape, query_dtype),  # output\n    )\n\ndef _dot_product_attention_bwd_abstract(\n    query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n    page_table_k, page_table_v, activation, fwd_output, grad_output, *,\n    scale, seed, dropout_rate, variadic_args, mask_type, layout, sliding_window_length):\n  query_dtype = dtypes.canonicalize_dtype(query.dtype)\n  key_dtype = dtypes.canonicalize_dtype(key.dtype)\n  value_dtype = dtypes.canonicalize_dtype(value.dtype)\n\n  _, has_dbias = variadic_args\n  if has_dbias:\n    # cuDNN supports bias for this case\n    bias_dtype = dtypes.canonicalize_dtype(bias.dtype)\n    return (\n      core.ShapedArray(\n          query.shape, query_dtype\n      ),  # grad query\n      core.ShapedArray(\n          key.shape, key_dtype\n      ),  # grad key\n      core.ShapedArray(\n          value.shape, value_dtype\n      ),  # grad value\n      core.ShapedArray(\n          bias.shape, bias_dtype\n      ),  # grad bias\n    )\n  else:\n    return (\n      core.ShapedArray(\n          query.shape, query_dtype\n      ),  # grad query\n      core.ShapedArray(\n          key.shape, key_dtype\n      ),  # grad key\n      core.ShapedArray(\n          value.shape, value_dtype\n      ),  # grad value\n    )\n\ndef _dot_product_attention_fwd_cuda_lowering(\n    ctx, query, key, value, bias, q_seqlen, kv_seqlen, q_offsets,\n    kv_offsets, page_table_k, page_table_v, scale, seed, dropout_rate,\n    variadic_args, mask_type, layout, sliding_window_length, is_training):\n  query_type = ir.RankedTensorType(query.type)\n  query_shape = query_type.shape\n  key_type = ir.RankedTensorType(key.type)\n  key_shape = key_type.shape\n\n  if layout == AttentionLayout.BNTH.value:\n    B, N, T, H = query_shape\n    _, _, S, _ = key_shape\n    output_layout = (3, 2, 1, 0)\n    output_transpose_perm = mlir.dense_int_array((0, 1, 2, 3))\n  else:\n    B, T, N, H = query_shape\n    _, S, _, _ = key_shape\n    output_layout = (3, 1, 2, 0)\n    output_transpose_perm = mlir.dense_int_array((0, 2, 1, 3))\n\n  max_seg_per_batch = get_max_seg_per_batch(ir.RankedTensorType(q_offsets.type))\n  is_paged_attention = check_is_paged_attention(ir.RankedTensorType(page_table_k.type))\n\n  output_shape = (B, N, T, H)\n  softmax_stat_shape = (B * max_seg_per_batch, N, T)\n  workspace_shape = (0,)\n  workspace_type = ir.IntegerType.get_unsigned(8)\n\n  has_bias, _ = variadic_args\n  backend_config = create_dot_product_attention_backend_config(\n      B, N, T, S, query_type.element_type, scale, seed, dropout_rate,\n      mask_type, layout, sliding_window_length, max_seg_per_batch,\n      is_paged_attention, is_bwd=False)\n  # {Q, K, V, bias*, q_seqlen*, kv_seqlen*,  q_offsets*, kv_offsets*}}\n  # {output, activation*, workspace}\n  has_dropout = dropout_rate > 0\n  operands = [query, key, value]\n  if has_bias:\n    operands.append(bias)\n  if has_padding(mask_type) or max_seg_per_batch > 1 or is_paged_attention:\n    operands.append(q_seqlen)\n    operands.append(kv_seqlen)\n  if max_seg_per_batch > 1:\n    operands.append(q_offsets)\n    operands.append(kv_offsets)\n  if is_paged_attention:\n    operands.append(page_table_k)\n    operands.append(page_table_v)\n\n  custom_call_name = get_custom_call_name(has_bias, has_dropout, False)\n\n  if is_training:\n    result_types = [\n      ir.RankedTensorType.get(output_shape, query_type.element_type),\n      ir.RankedTensorType.get(softmax_stat_shape, ir.F32Type.get()),\n      ir.RankedTensorType.get(workspace_shape, workspace_type),\n    ]\n    result_layouts = [output_layout] + default_layouts(softmax_stat_shape, workspace_shape)\n  else:\n    result_types = [\n      ir.RankedTensorType.get(output_shape, query_type.element_type),\n      ir.RankedTensorType.get(workspace_shape, workspace_type)\n    ]\n    result_layouts = [output_layout] + default_layouts(workspace_shape)\n  # create custom call here\n  out = mlir.custom_call(\n    custom_call_name,\n    result_types=result_types,\n    operands=operands,\n    backend_config=backend_config,\n    operand_layouts=default_layouts(\n      *[ir.RankedTensorType(operand.type).shape for operand in operands]),\n    result_layouts=result_layouts,\n  )\n  # drop workspace memory\n  # output should be (B, T, N, H) instead of (B, N, T, H)\n  if is_training:\n    return [hlo.transpose(out.results[0], output_transpose_perm), out.results[1]]\n  else:\n    return [hlo.transpose(out.results[0], output_transpose_perm)]\n\ndef _dot_product_attention_bwd_cuda_lowering(\n    ctx, query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n    page_table_k, page_table_v, activation, fwd_output, grad_output,\n    scale, seed, dropout_rate, variadic_args, mask_type, layout, sliding_window_length):\n  query_type = ir.RankedTensorType(query.type)\n  query_shape = query_type.shape\n  key_type = ir.RankedTensorType(key.type)\n  key_shape = key_type.shape\n  value_type = ir.RankedTensorType(value.type)\n\n  if layout == AttentionLayout.BNTH.value:\n    B, q_N, T, H = query_shape\n    _, k_N, S, _ = key_shape\n    grad_layout = (3, 2, 1, 0)\n    grad_transpose_perm = mlir.dense_int_array((0, 1, 2, 3))\n  else:\n    B, T, q_N, H = query_shape\n    _, S, k_N, _ = key_shape\n    grad_layout = (3, 1, 2, 0)\n    grad_transpose_perm = mlir.dense_int_array((0, 2, 1, 3))\n\n  workspace_shape = (0,)\n  workspace_type = ir.IntegerType.get_unsigned(8)\n\n  grad_query_shape = (B, q_N, T, H)\n  grad_key_shape = (B, k_N, S, H)\n  grad_value_shape = (B, k_N, S, H)\n\n  has_bias, has_dbias = variadic_args\n  max_seg_per_batch = get_max_seg_per_batch(ir.RankedTensorType(q_offsets.type))\n  backend_config = create_dot_product_attention_backend_config(\n      B, q_N, T, S, query_type.element_type, scale, seed, dropout_rate,\n      mask_type, layout, sliding_window_length, max_seg_per_batch,\n      False, is_bwd=True)\n  # {Q, K, V, activation, dO, bias*, O, q_seqlen*, kv_seqlen*,\n  #  q_offsets*, kv_offsets*}\n  # {dQ, dK, dV, dbias*, workspace}\n  has_dropout = dropout_rate > 0\n  # create operands\n  operands = [query, key, value, activation, grad_output]\n  if has_bias:\n    # flash attention requires bias in the bwd for remat\n    operands.append(bias)\n  operands.append(fwd_output)\n  if has_padding(mask_type) or max_seg_per_batch > 1:\n    operands.append(q_seqlen)\n    operands.append(kv_seqlen)\n  if max_seg_per_batch > 1:\n    operands.append(q_offsets)\n    operands.append(kv_offsets)\n  # get custom call name\n  custom_call_name = get_custom_call_name(has_bias, has_dropout, True)\n\n  # create output types and layouts\n  # grad_query, grad_key, grad_value\n  result_types = [\n    ir.RankedTensorType.get(grad_query_shape, query_type.element_type),\n    ir.RankedTensorType.get(grad_key_shape, key_type.element_type),\n    ir.RankedTensorType.get(grad_value_shape, value_type.element_type),\n  ]\n  result_layouts = [grad_layout, grad_layout, grad_layout]\n  bias_type = ir.RankedTensorType(bias.type)\n  bias_shape = bias_type.shape\n  if has_dbias:\n    # cuDNN supports bias for this case\n    result_types.append(\n      ir.RankedTensorType.get(bias_shape, bias_type.element_type))\n    result_layouts = result_layouts + default_layouts(bias_shape)\n  # workspace\n  result_types.append(ir.RankedTensorType.get(workspace_shape, workspace_type))\n  result_layouts = result_layouts + default_layouts(workspace_shape)\n  out = mlir.custom_call(\n    custom_call_name,\n    result_types=result_types,\n    operands=operands,\n    backend_config=backend_config,\n    operand_layouts=default_layouts(\n      *[ir.RankedTensorType(operand.type).shape for operand in operands]),\n    result_layouts=result_layouts,\n  )\n  dqkv = (hlo.transpose(out.results[0], grad_transpose_perm),\n          hlo.transpose(out.results[1], grad_transpose_perm),\n          hlo.transpose(out.results[2], grad_transpose_perm))\n  # Only keep dQ, dK, dV and dBias here\n  if has_dbias:\n    return dqkv + (out.results[3],)\n  else:\n    return dqkv\n\n# batcher\ndef _check_valid_batch_dims(bdims):\n  for dim in bdims:\n    if dim not in [0, None]:\n      raise NotImplementedError(\n        f""Currently only support batch_dim in [0, None], but got {dim=}"")\n\ndef _dot_product_attention_fwd_batcher(\n    batched_args, batch_dims, *, scale, seed, dropout_rate, variadic_args,\n    mask_type, layout, sliding_window_length, is_training):\n  _check_valid_batch_dims(batch_dims)\n  query, key, value, bias, q_seqlen, kv_seqlen, \\n    q_offsets, kv_offsets, page_table_k, page_table_v = batched_args\n  query_bdim = batch_dims[0]\n  if is_training:\n    out_bdims = query_bdim, query_bdim\n  else:\n    out_bdims = (query_bdim,)\n\n  if layout == AttentionLayout.BNTH.value:\n    *Bs, N, T, _ = query.shape\n    *_, _, S, _ = key.shape\n  else:\n    *Bs, T, N, _ = query.shape\n    *_, S, _, _ = key.shape\n  B = math.prod(Bs)\n  has_bias, _ = variadic_args\n  original_shape = query.shape\n  # reshape to 4D shape\n  query = jnp.reshape(query, (B,) + query.shape[-3:])\n  key = jnp.reshape(key, (B,) + key.shape[-3:])\n  value = jnp.reshape(value, (B,) + key.shape[-3:])\n  if has_bias and batch_dims[3] is not None:\n    bias = jnp.reshape(bias, (B, N, T, S))\n  if has_padding(mask_type):\n    q_seqlen = jnp.reshape(q_seqlen, (B, ))\n    kv_seqlen = jnp.reshape(kv_seqlen, (B, ))\n\n  outputs = _dot_product_attention_fwd_p_wrapper.bind(\n      query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n      page_table_k, page_table_v, scale=scale, seed=seed, dropout_rate=dropout_rate,\n      variadic_args=variadic_args, mask_type=mask_type, layout=layout,\n      sliding_window_length=sliding_window_length, is_training=is_training)\n\n  # reshape to original shape\n  output = outputs[0]\n  output = jnp.reshape(output, original_shape)\n  if is_training:\n    activation = outputs[1]\n    activation = jnp.reshape(activation, (*Bs, N, T))\n    return (output, activation), out_bdims\n  else:\n    return (output,), out_bdims\n\ndef _dot_product_attention_bwd_batcher(\n     batched_args, batch_dims, *, scale, seed, dropout_rate, variadic_args,\n     mask_type, layout, sliding_window_length):\n  _check_valid_batch_dims(batch_dims)\n  query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets, \\n    page_table_k, page_table_v, activation, fwd_output, grad_output = batched_args\n  query_bdim = batch_dims[0]\n  out_bdims = query_bdim, query_bdim, query_bdim\n\n  if layout == AttentionLayout.BNTH.value:\n    *Bs, N, T, _ = query.shape\n    *_, _, S, _ = key.shape\n  else:\n    *Bs, T, N, _ = query.shape\n    *_, S, _, _ = key.shape\n  B = math.prod(Bs)\n  has_bias, has_dbias = variadic_args\n  # Reset the has_dbias if the combined batch size is not 1, because cuDNN only\n  # supports dbias with a single batch. In this case, an all-zero dbias will be\n  # appended instead.\n  if B > 1:\n    variadic_args = (has_bias, False)\n  original_query_shape = query.shape\n  original_key_shape = key.shape\n  original_value_shape = value.shape\n  original_bias_shape = bias.shape if has_bias else None\n  # reshape to 4D shape\n  query = jnp.reshape(query, (B,) + query.shape[-3:])\n  key = jnp.reshape(key, (B,) + key.shape[-3:])\n  value = jnp.reshape(value, (B,) + key.shape[-3:])\n  if has_bias and batch_dims[3] is not None:\n    bias = jnp.reshape(bias, (B, N, T, S))\n  if has_padding(mask_type):\n    q_seqlen = jnp.reshape(q_seqlen, (B, ))\n    kv_seqlen = jnp.reshape(kv_seqlen, (B, ))\n\n  activation = jnp.reshape(activation, (B, N, T))\n  fwd_output = jnp.reshape(fwd_output, (B,) + query.shape[-3:])\n  grad_output = jnp.reshape(grad_output, (B,) + query.shape[-3:])\n\n  grads = _dot_product_attention_bwd_p_wrapper.bind(\n      query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n      page_table_k, page_table_v, activation, fwd_output, grad_output,\n      scale=scale, seed=seed, dropout_rate=dropout_rate, variadic_args=variadic_args,\n      mask_type=mask_type, layout=layout,\n      sliding_window_length=sliding_window_length,\n  )\n\n  # reshape to original shape\n  grads[0] = jnp.reshape(grads[0], original_query_shape)\n  grads[1] = jnp.reshape(grads[1], original_key_shape)\n  grads[2] = jnp.reshape(grads[2], original_value_shape)\n  if has_dbias:\n    assert has_bias\n    if variadic_args[1]:\n      grads[3] = jnp.reshape(grads[3], original_bias_shape)\n    else:\n      grads.append(jnp.zeros(original_bias_shape, bias.dtype))\n    out_bdims += (batch_dims[3],)\n  return grads, out_bdims\n\n# custom partitioning\ndef _get_padded_spec(arg_info):\n  spec = None if arg_info.sharding is None else arg_info.sharding.spec\n  ndim = arg_info.ndim\n  if spec is None:\n    return (None,) * ndim\n  assert len(spec) <= ndim\n  return spec + (None,) * (ndim - len(spec))\n\ndef _check_qkv_bias_mask_spec(\n    query_spec, key_spec, value_spec, bias_spec, layout):\n  # check qkv spec\n  if not query_spec == key_spec == value_spec:\n    raise ValueError(""Query, key and value should have same sharding."")\n  if layout == AttentionLayout.BNTH.value:\n    *batch_spec, num_head_spec, q_seq_spec, head_spec = query_spec\n  else:\n    *batch_spec, q_seq_spec, num_head_spec, head_spec = query_spec\n  if q_seq_spec is not None:\n    raise ValueError(""Sharding on sequence dim is not allowed."")\n  if head_spec is not None:\n    raise ValueError(""Sharding on head dim is not allowed."")\n  # check bias spec\n  if bias_spec:\n    *bias_batch_spec, bias_num_head_spec, bias_q_seq_spec, bias_kv_seq_spec = bias_spec\n    if any(bias_batch_spec) and bias_batch_spec != batch_spec or \\n      bias_num_head_spec is not None and bias_num_head_spec != num_head_spec:\n      raise ValueError(\n        ""Query and bias should have same sharding on batch and num_head dim."")\n    if bias_q_seq_spec is not None or bias_kv_seq_spec is not None:\n      raise ValueError(""Sharding on bias sequence dim is not allowed."")\n\n\n# fwd custom partition\ndef _infer_fwd_output_sharding(mesh, arg_shapes, variadic_args,is_training, layout):\n  # only sharding on batch and num_head dim is allowed\n  # (*batch, q_seq, num_head, head)\n  query_spec = _get_padded_spec(arg_shapes[0])\n  # (*batch, kv_seq, num_head, head)\n  key_spec = _get_padded_spec(arg_shapes[1])\n  value_spec = _get_padded_spec(arg_shapes[2])\n  has_bias, _ = variadic_args\n  bias_spec = _get_padded_spec(arg_shapes[3]) if has_bias else None\n\n  _check_qkv_bias_mask_spec(\n    query_spec, key_spec, value_spec, bias_spec, layout)\n  # keep out sharding same as query sharding since they have same shape\n  out_sharding = NamedSharding(mesh, PartitionSpec(*query_spec))\n  if is_training:\n    # activation sharding\n    *batch_spec, q_seq_spec, num_head_spec, _ = query_spec\n    activation_sharding = NamedSharding(\n      mesh, PartitionSpec(*batch_spec, num_head_spec, q_seq_spec, None))\n    return [out_sharding, activation_sharding]\n  return [out_sharding]\n\n_dot_product_attention_fwd_lower = custom_partitioning(\n    _dot_product_attention_fwd_impl, static_argnums=(10, 11, 12, 13, 14, 15, 16, 17))\n\ndef _dot_product_attention_fwd_infer_sharding_from_operands(\n    scale, seed, dropout_rate, variadic_args, mask_type, layout, sliding_window_length,\n    is_training, mesh, arg_shapes, result_shape):\n  return _infer_fwd_output_sharding(mesh, arg_shapes, variadic_args, is_training, layout)\n\ndef _dot_product_attention_fwd_partition(\n    scale, seed, dropout_rate, variadic_args, mask_type, layout, sliding_window_length,\n    is_training, mesh, arg_shapes, result_shape):\n  # args sharding\n  arg_shardings = tuple(arg_i.sharding for arg_i in arg_shapes)\n  out_shardings = _infer_fwd_output_sharding(\n    mesh, arg_shapes, variadic_args, is_training, layout)\n  impl = functools.partial(\n      _dot_product_attention_fwd_impl,\n      scale=scale,\n      seed=seed,\n      dropout_rate=dropout_rate,\n      variadic_args=variadic_args,\n      mask_type=mask_type,\n      layout=layout,\n      sliding_window_length=sliding_window_length,\n      is_training=is_training,\n  )\n  return mesh, impl, out_shardings, arg_shardings\n\n# bwd custom partition\ndef _infer_bwd_output_sharding(mesh, arg_shapes, layout, variadic_args):\n  # (*batch, q_seq, num_head, head)\n  query_spec = _get_padded_spec(arg_shapes[0])\n  # (*batch, kv_seq, num_head, head)\n  key_spec = _get_padded_spec(arg_shapes[1])\n  value_spec = _get_padded_spec(arg_shapes[2])\n  has_bias, has_dbias = variadic_args\n  bias_spec = _get_padded_spec(arg_shapes[3]) if has_bias else None\n  _check_qkv_bias_mask_spec(\n    query_spec, key_spec, value_spec, bias_spec, layout)\n  # keep grad query sharding same as query sharding\n  grad_query_sharding = NamedSharding(mesh, PartitionSpec(*query_spec))\n  grad_key_sharding = NamedSharding(mesh, PartitionSpec(*key_spec))\n  grad_value_sharding = NamedSharding(mesh, PartitionSpec(*key_spec))\n  out_shardings = [grad_query_sharding, grad_key_sharding, grad_value_sharding]\n  if has_dbias:\n    grad_bias_sharding = NamedSharding(mesh, PartitionSpec(*bias_spec))\n    out_shardings = out_shardings + [grad_bias_sharding]\n  return out_shardings\n\n_dot_product_attention_bwd_lower = custom_partitioning(\n    _dot_product_attention_bwd_impl, static_argnums=(13, 14, 15, 16, 17, 18, 19)\n)\n\ndef _dot_product_attention_bwd_infer_sharding_from_operands(\n    scale, seed, dropout_rate, variadic_args, mask_type, layout,\n    sliding_window_length, mesh, arg_shapes, result_shape):\n  return _infer_bwd_output_sharding(mesh, arg_shapes, layout, variadic_args)\n\ndef _dot_product_attention_bwd_partition(\n    scale, seed, dropout_rate, variadic_args, mask_type, layout,\n    sliding_window_length, mesh, arg_shapes, result_shape):\n  out_shardings = _infer_bwd_output_sharding(mesh, arg_shapes, layout, variadic_args)\n  # args sharding\n  arg_shardings = tuple(arg_i.sharding for arg_i in arg_shapes)\n  def sharded_impl(*args):\n    impl = functools.partial(\n      _dot_product_attention_bwd_impl,\n      scale=scale,\n      seed=seed,\n      dropout_rate=dropout_rate,\n      variadic_args=variadic_args,\n      mask_type=mask_type,\n      layout=layout,\n      sliding_window_length=sliding_window_length,\n    )\n    grads = impl(*args)\n    _, has_dbias = variadic_args\n    if has_dbias:\n      query_spec = arg_shardings[0].spec\n      batch_spec = query_spec[0]\n      local_dbias = grads[3]\n      global_dbias = jax.lax.psum(local_dbias, batch_spec)\n      grads = grads[:3] + [global_dbias]\n    return grads\n  return mesh, sharded_impl, out_shardings, arg_shardings\n\n# Create dot_product_attention_fwd_p for forward operation.\n_dot_product_attention_fwd_p = core.Primitive(""dot_product_attention_fwd"")\n_dot_product_attention_fwd_p.multiple_results = True\n_dot_product_attention_fwd_p.def_impl(\n    functools.partial(dispatch.apply_primitive, _dot_product_attention_fwd_p)\n)\n_dot_product_attention_fwd_p.def_abstract_eval(\n    _dot_product_attention_fwd_abstract\n)\n\nmlir.register_lowering(\n  _dot_product_attention_fwd_p,\n  _dot_product_attention_fwd_cuda_lowering,\n  platform=""cuda"",\n)\n\n_dot_product_attention_fwd_p_wrapper = core.Primitive(\n    ""dot_product_attention_fwd_wrapper""\n)\n_dot_product_attention_fwd_p_wrapper.multiple_results = True\n_dot_product_attention_fwd_p_wrapper.def_impl(_dot_product_attention_fwd_impl)\n_dot_product_attention_fwd_p_wrapper.def_abstract_eval(\n    _dot_product_attention_fwd_abstract\n)\n\n# Create dot_product_attention_bwd_p for backward operation.\n_dot_product_attention_bwd_p = core.Primitive(""dot_product_attention_bwd"")\n_dot_product_attention_bwd_p.multiple_results = True\n_dot_product_attention_bwd_p.def_impl(\n    functools.partial(dispatch.apply_primitive, _dot_product_attention_bwd_p)\n)\n_dot_product_attention_bwd_p.def_abstract_eval(\n    _dot_product_attention_bwd_abstract\n)\n\nmlir.register_lowering(\n  _dot_product_attention_bwd_p,\n  _dot_product_attention_bwd_cuda_lowering,\n  platform=""cuda"",\n)\n\n_dot_product_attention_bwd_p_wrapper = core.Primitive(\n    ""dot_product_attention_bwd_wrapper""\n)\n_dot_product_attention_bwd_p_wrapper.multiple_results = True\n_dot_product_attention_bwd_p_wrapper.def_impl(_dot_product_attention_bwd_impl)\n_dot_product_attention_bwd_p_wrapper.def_abstract_eval(\n    _dot_product_attention_bwd_abstract\n)\n\nbatching.primitive_batchers[\n    _dot_product_attention_fwd_p_wrapper\n] = _dot_product_attention_fwd_batcher\nbatching.primitive_batchers[\n    _dot_product_attention_bwd_p_wrapper\n] = _dot_product_attention_bwd_batcher\n\ndef not_implemented_sharding_rule(*args, **kwargs):\n  return NotImplementedError(""Sharding rule not implemented."")\n\n_dot_product_attention_fwd_lower.def_partition(\n  infer_sharding_from_operands=_dot_product_attention_fwd_infer_sharding_from_operands,\n  partition=_dot_product_attention_fwd_partition,\n  sharding_rule=not_implemented_sharding_rule)\n\nmlir.register_lowering(_dot_product_attention_fwd_p_wrapper,\n                        mlir.lower_fun(_dot_product_attention_fwd_lower, multiple_results=True))\n\n_dot_product_attention_bwd_lower.def_partition(\n  infer_sharding_from_operands=_dot_product_attention_bwd_infer_sharding_from_operands,\n  partition=_dot_product_attention_bwd_partition,\n  sharding_rule=not_implemented_sharding_rule)\n\nmlir.register_lowering(_dot_product_attention_bwd_p_wrapper,\n                        mlir.lower_fun(_dot_product_attention_bwd_lower, multiple_results=True))\n\ndispatch.prim_requires_devices_during_lowering.add(\n  _dot_product_attention_fwd_p\n)\ndispatch.prim_requires_devices_during_lowering.add(\n  _dot_product_attention_fwd_p_wrapper\n)\ndispatch.prim_requires_devices_during_lowering.add(\n  _dot_product_attention_bwd_p\n)\ndispatch.prim_requires_devices_during_lowering.add(\n  _dot_product_attention_bwd_p_wrapper\n)\n\n@functools.partial(jax.custom_vjp, nondiff_argnums=(10, 11, 12, 13, 14, 15, 16, 17, 18))\ndef _dot_product_attention(query: Array,\n                           key: Array,\n                           value: Array,\n                           bias: Array,\n                           q_seqlen: Array,\n                           kv_seqlen: Array,\n                           q_offsets: Array,\n                           kv_offsets: Array,\n                           page_table_k: Array,\n                           page_table_v: Array,\n                           scale: float,\n                           seed: int,\n                           dropout_rate: float,\n                           variadic_args: tuple[bool, ...],\n                           mask_type: bool,\n                           layout: int,\n                           sliding_window_length: int | None,\n                           cudnn_version: int,\n                           return_residual: bool):\n  output = _dot_product_attention_fwd(\n      query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n      page_table_k, page_table_v, scale=scale, seed=seed, dropout_rate=dropout_rate,\n      variadic_args=variadic_args, mask_type=mask_type, layout=layout,\n      sliding_window_length=sliding_window_length,\n      cudnn_version=cudnn_version, return_residual=return_residual)\n  return output\n\n_dot_product_attention.defvjp(\n    _dot_product_attention_fwd_rule, _dot_product_attention_bwd_rule\n)\n\nfp8_params_keys = [\n    'amax_dQ', 'amax_dK', 'amax_dV', 'amax_dP', # place holder for bwd output\n    'descale_q', 'descale_k', 'descale_v', 'descale_s',\n    'scale_s', 'scale_o', 'descale_o', 'descale_dO',\n    'descale_dP', 'scale_dQ', 'scale_dK', 'scale_dV',\n    'scale_dP'\n]\n\nfp8_params_keys_fwd = [\n    'descale_q', 'descale_k', 'descale_v', 'descale_s', 'scale_s', 'scale_o'\n]\nfp8_params_keys_bwd = [\n    'descale_q', 'descale_k', 'descale_v', 'descale_o', 'descale_dO', 'descale_s',\n    'descale_dP', 'scale_s', 'scale_dQ', 'scale_dK', 'scale_dV', 'scale_dP',\n]\nparams_from_keys = lambda params, keys: [params[key] for key in keys]\n\ndef check_fp8_params(params):\n  # Check if all required keys are present\n  missing_keys = set(fp8_params_keys) - set(params)\n  if missing_keys:\n    raise ValueError(f""The following keys are missing from fp8_params: {', '.join(missing_keys)}"")\n\ncheck_is_flash_attention_fp8 = functools.partial(\n    check_is_flash_attention,\n    has_bias=False,\n    is_fp8=True\n)\n\ndef _dot_product_attention_fp8_fwd(\n    query, key, value,\n    fp8_params_fwd,\n    scale, use_causal_mask, layout, cudnn_version):\n  check_is_flash_attention_fp8(\n      query, key, layout, cudnn_version, is_training=False)\n  descale_q, descale_k, descale_v, descale_s, scale_s, scale_o = fp8_params_fwd\n  outputs = _dot_product_attention_fp8_fwd_p_wrapper.bind(\n      query, key, value,\n      descale_q, descale_k, descale_v, descale_s,\n      scale_s, scale_o,\n      scale=scale, use_causal_mask=use_causal_mask, layout=layout, is_training=False)\n  return outputs\n\ndef _dot_product_attention_fp8_fwd_rule(\n    query, key, value,\n    fp8_params,\n    scale, use_causal_mask, layout, cudnn_version):\n  check_is_flash_attention_fp8(\n      query, key, layout, cudnn_version, is_training=True)\n\n  outputs = _dot_product_attention_fp8_fwd_p_wrapper.bind(\n      query, key, value, *params_from_keys(fp8_params, fp8_params_keys_fwd),\n      scale=scale, use_causal_mask=use_causal_mask, layout=layout, is_training=True)\n  res = (query, key, value, outputs[3], outputs[0], params_from_keys(fp8_params, fp8_params_keys_bwd))\n  return (outputs[0], outputs[1], outputs[2]), res\n\ndef _dot_product_attention_fp8_bwd_rule(\n    scale, use_causal_mask, layout, cudnn_version, res, g):\n  (query, key, value, activation, fwd_output, aux_params) = res\n  grad_output = g[0]\n  grads = _dot_product_attention_fp8_bwd_p_wrapper.bind(\n    query,\n    key,\n    value,\n    fwd_output,\n    grad_output,\n    activation,\n    *aux_params,\n    scale=scale,\n    use_causal_mask=use_causal_mask,\n    layout=layout,\n    )\n\n  fp8_params_grads = dict.fromkeys(fp8_params_keys)\n  keys_to_grad_indices = ['amax_dQ', 'amax_dK', 'amax_dV', 'amax_dP']\n  # grads structure: (dQ, dK, dV, amax_dq, amax_dk, amax_dv, amax_dp)\n  for i, key in enumerate(keys_to_grad_indices, start=3):\n    fp8_params_grads[key] = grads[i]\n\n  return (grads[0], grads[1], grads[2], fp8_params_grads)\n\ndef _dot_product_attention_fp8_fwd_impl(\n    query, key, value,\n    descale_q, descale_k, descale_v, descale_s, scale_s, scale_o,\n    scale, use_causal_mask, layout, is_training):\n  outputs = _dot_product_attention_fp8_fwd_p.bind(\n      query,\n      key,\n      value,\n      descale_q,\n      descale_k,\n      descale_v,\n      descale_s,\n      scale_s,\n      scale_o,\n      scale=scale,\n      use_causal_mask=use_causal_mask,\n      layout=layout,\n      is_training=is_training,\n  )\n  return outputs\n\ndef _dot_product_attention_fp8_bwd_impl(\n    query, key, value, fwd_output, grad_output, activation,\n    descale_q, descale_k, descale_v, descale_o, descale_dO, descale_s,\n    descale_dP, scale_s, scale_dQ, scale_dK, scale_dV, scale_dP,\n    scale, use_causal_mask, layout):\n  grads = _dot_product_attention_fp8_bwd_p.bind(\n      query, key, value, fwd_output, grad_output, activation,\n      descale_q, descale_k, descale_v, descale_o, descale_dO, descale_s,\n      descale_dP, scale_s, scale_dQ, scale_dK, scale_dV, scale_dP,\n      scale=scale, use_causal_mask=use_causal_mask, layout=layout)\n  return grads\n\n\ndef _dot_product_attention_fp8_fwd_abstract(\n    query, key, value,\n    descale_q, descale_k, descale_v, descale_s, scale_s, scale_o,\n    scale, use_causal_mask, layout, is_training):\n  query_dtype = dtypes.canonicalize_dtype(query.dtype)\n  if layout == AttentionLayout.BNTH.value:\n    B, N, T, _ = query.shape\n    _, _, S, _ = key.shape\n  else:\n    B, T, N, _ = query.shape\n    _, S, _, _ = key.shape\n  output_shape = query.shape\n  softmax_stat_shape = (B, N, T)\n\n  # output, amax_s, amax_o[, softmax_stat]\n  if is_training:\n    return (\n      core.ShapedArray(output_shape, query_dtype),\n      core.ShapedArray((1,1,1,1), jnp.float32),\n      core.ShapedArray((1,1,1,1), jnp.float32),\n      core.ShapedArray(softmax_stat_shape, jnp.float32),\n    )\n  else:\n    return (\n      core.ShapedArray(output_shape, query_dtype),\n      core.ShapedArray((1,1,1,1), jnp.float32),\n      core.ShapedArray((1,1,1,1), jnp.float32),\n    )\n\ndef _dot_product_attention_fp8_bwd_abstract(\n    query, key, value, fwd_output, grad_output, activation,\n    descale_q, descale_k, descale_v, descale_o, descale_dO, descale_s,\n    descale_dP, scale_s, scale_dQ, scale_dK, scale_dV, scale_dP,\n    scale, use_causal_mask, layout):\n  query_dtype = dtypes.canonicalize_dtype(query.dtype)\n  key_dtype = dtypes.canonicalize_dtype(key.dtype)\n  value_dtype = dtypes.canonicalize_dtype(value.dtype)\n\n  amax_shape = (1,1,1,1)\n\n  return (\n    core.ShapedArray(query.shape, query_dtype),\n    core.ShapedArray(key.shape, key_dtype),\n    core.ShapedArray(value.shape, value_dtype),\n    core.ShapedArray(amax_shape, jnp.float32),\n    core.ShapedArray(amax_shape, jnp.float32),\n    core.ShapedArray(amax_shape, jnp.float32),\n    core.ShapedArray(amax_shape, jnp.float32),\n  )\n\ndef _dot_product_attention_fp8_fwd_cuda_lowering(\n    ctx, query, key, value,\n    descale_q, descale_k, descale_v, descale_s, scale_s, scale_o,\n    scale, use_causal_mask, layout, is_training):\n  query_type = ir.RankedTensorType(query.type)\n  query_shape = query_type.shape\n  key_type = ir.RankedTensorType(key.type)\n  key_shape = key_type.shape\n\n  if layout == AttentionLayout.BNTH.value:\n    B, N, T, H = query_shape\n    _, _, S, _ = key_shape\n    output_layout = (3, 2, 1, 0)\n    output_transpose_perm = mlir.dense_int_array((0, 1, 2, 3))\n  else:\n    B, T, N, H = query_shape\n    _, S, _, _ = key_shape\n    output_layout = (3, 1, 2, 0)\n    output_transpose_perm = mlir.dense_int_array((0, 2, 1, 3))\n\n  output_shape = (B, N, T, H)\n  softmax_stat_shape = (B, N, T)\n  workspace_shape = (0,)\n  amax_shape = (1,1,1,1)\n  workspace_type = ir.IntegerType.get_unsigned(8)\n  mask_type = MaskType.CAUSAL if use_causal_mask else MaskType.NO_MASK\n  backend_config = create_dot_product_attention_fp8_backend_config(\n      B, N, T, S, ir.BF16Type.get(),  # query_type.element_type,\n      scale, mask_type, layout, is_bwd=False,\n  )\n\n  operands = [query, key, value, descale_q, descale_k, descale_v, descale_s, scale_s, scale_o]\n  custom_call_name = get_fp8_custom_call_name(is_bwd=False)\n\n  if is_training:\n    result_types = [\n      ir.RankedTensorType.get(output_shape, query_type.element_type),\n      ir.RankedTensorType.get((1,1,1,1), ir.F32Type.get()),\n      ir.RankedTensorType.get((1,1,1,1), ir.F32Type.get()),\n      ir.RankedTensorType.get(softmax_stat_shape, ir.F32Type.get()),\n      ir.RankedTensorType.get(workspace_shape, workspace_type),\n    ]\n    result_layouts = [output_layout] + default_layouts(amax_shape, amax_shape, softmax_stat_shape, workspace_shape)\n  else:\n    result_types = [\n      ir.RankedTensorType.get(output_shape, query_type.element_type),\n      ir.RankedTensorType.get((1,1,1,1), ir.F32Type.get()),\n      ir.RankedTensorType.get((1,1,1,1), ir.F32Type.get()),\n      ir.RankedTensorType.get(workspace_shape, workspace_type)\n    ]\n    result_layouts = [output_layout] + default_layouts(amax_shape, amax_shape, workspace_shape)\n\n  operand_shapes = [ir.RankedTensorType(operand.type).shape for operand in operands[:3]]\n  operand_shapes += [[1, 1, 1, 1]] * 6\n  operand_layouts = default_layouts(*operand_shapes)\n  out = mlir.custom_call(\n    custom_call_name,\n    result_types=result_types,\n    operands=operands,\n    backend_config=backend_config,\n    operand_layouts=operand_layouts,\n    result_layouts=result_layouts,\n  )\n\n  if is_training:\n    return [hlo.transpose(out.results[0], output_transpose_perm), out.results[1], out.results[2], out.results[3]]\n  else:\n    return [hlo.transpose(out.results[0], output_transpose_perm), out.results[1], out.results[2]]\n\n\n\ndef _dot_product_attention_fp8_bwd_cuda_lowering(\n    ctx, query, key, value, fwd_output, grad_output, activation,\n    descale_q, descale_k, descale_v, descale_o, descale_dO, descale_s,\n    descale_dP, scale_s, scale_dQ, scale_dK, scale_dV, scale_dP, scale,\n    use_causal_mask, layout):\n  query_type = ir.RankedTensorType(query.type)\n  query_shape = query_type.shape\n  key_type = ir.RankedTensorType(key.type)\n  key_shape = key_type.shape\n  value_type = ir.RankedTensorType(value.type)\n\n  if layout == AttentionLayout.BNTH.value:\n    B, q_N, T, H = query_shape\n    _, k_N, S, _ = key_shape\n    grad_layout = (3, 2, 1, 0)\n    grad_transpose_perm = mlir.dense_int_array((0, 1, 2, 3))\n  else:\n    B, T, q_N, H = query_shape\n    _, S, k_N, _ = key_shape\n    grad_layout = (3, 1, 2, 0)\n    grad_transpose_perm = mlir.dense_int_array((0, 2, 1, 3))\n\n  workspace_shape = (0,)\n  workspace_type = ir.IntegerType.get_unsigned(8)\n  amax_shape = (1,1,1,1)\n\n  grad_query_shape = (B, q_N, T, H)\n  grad_key_shape = (B, k_N, S, H)\n  grad_value_shape = (B, k_N, S, H)\n  mask_type = MaskType.CAUSAL if use_causal_mask else MaskType.NO_MASK\n\n  backend_config = create_dot_product_attention_fp8_backend_config(\n      B, q_N, T, S, ir.BF16Type.get(),\n      scale, mask_type, layout, is_bwd=True,\n  )\n\n  operands = [\n    query,\n    key,\n    value,\n    fwd_output,\n    grad_output,\n    activation,\n    descale_q,\n    descale_k,\n    descale_v,\n    descale_o,\n    descale_dO,\n    descale_s,\n    descale_dP,\n    scale_s,\n    scale_dQ,\n    scale_dK,\n    scale_dV,\n    scale_dP,\n  ]\n\n  custom_call_name = get_fp8_custom_call_name(is_bwd=True)\n\n  result_types = [\n    ir.RankedTensorType.get(grad_query_shape, query_type.element_type),\n    ir.RankedTensorType.get(grad_key_shape, key_type.element_type),\n    ir.RankedTensorType.get(grad_value_shape, value_type.element_type),\n    ir.RankedTensorType.get(amax_shape, ir.F32Type.get()),\n    ir.RankedTensorType.get(amax_shape, ir.F32Type.get()),\n    ir.RankedTensorType.get(amax_shape, ir.F32Type.get()),\n    ir.RankedTensorType.get(amax_shape, ir.F32Type.get()),\n  ]\n  result_layouts = [grad_layout, grad_layout, grad_layout] + default_layouts(amax_shape, amax_shape, amax_shape, amax_shape)\n\n  result_types.append(ir.RankedTensorType.get(workspace_shape, workspace_type))\n  result_layouts = result_layouts + default_layouts(workspace_shape)\n  out = mlir.custom_call(\n    custom_call_name,\n    result_types=result_types,\n    operands=operands,\n    backend_config=backend_config,\n    operand_layouts=default_layouts(\n      *[ir.RankedTensorType(operand.type).shape for operand in operands]),\n    result_layouts=result_layouts,\n  )\n  dqkv_amaxs = (hlo.transpose(out.results[0], grad_transpose_perm),\n          hlo.transpose(out.results[1], grad_transpose_perm),\n          hlo.transpose(out.results[2], grad_transpose_perm),\n          out.results[3], out.results[4], out.results[5], out.results[6])\n  # Only keep dQ, dK, dV, amax_dQ, amax_dK, amax_dV, amax_dP here\n  return dqkv_amaxs\n\ndef _dot_product_attention_fp8_fwd_batcher(\n    batched_args, batch_dims, *, scale, use_causal_mask, layout, is_training):\n  _check_valid_batch_dims(batch_dims)\n  query, key, value,\\n    descale_q, descale_k, descale_v, descale_s, scale_s, scale_o, = batched_args\n  query_bdim = batch_dims[0]\n  if is_training:\n    out_bdims = query_bdim, query_bdim\n  else:\n    out_bdims = (query_bdim,)\n\n  if layout == AttentionLayout.BNTH.value:\n    *Bs, N, T, _ = query.shape\n    *_, _, S, _ = key.shape\n  else:\n    *Bs, T, N, _ = query.shape\n    *_, S, _, _ = key.shape\n  B = math.prod(Bs)\n\n  # reshape to 4D shape\n  query = jnp.reshape(query, (B,) + query.shape[-3:])\n  key = jnp.reshape(key, (B,) + key.shape[-3:])\n  value = jnp.reshape(value, (B,) + key.shape[-3:])\n\n  outputs = _dot_product_attention_fp8_fwd_p_wrapper.bind(\n      query, key, value, descale_q, descale_k, descale_v, descale_s, scale_s, scale_o,\n      scale=scale, use_causal_mask=use_causal_mask, layout=layout, is_training=is_training)\n\n  # reshape to original shape\n  output, amax_s, amax_o = outputs[0], outputs[1], outputs[2]\n  output = jnp.reshape(output, query.shape)\n  if is_training:\n    activation = outputs[3]\n    activation = jnp.reshape(activation, (*Bs, N, T))\n    return (output, amax_s, amax_o, activation), out_bdims\n  else:\n    return (output, amax_s, amax_o), out_bdims\n\ndef _dot_product_attention_fp8_bwd_batcher(\n    batched_args, batch_dims, *, scale, use_causal_mask, layout):\n  _check_valid_batch_dims(batch_dims)\n  query, key, value, fwd_output, grad_output, activation,\\n    descale_q, descale_k, descale_v, descale_o, descale_dO, descale_s, descale_dP,\\n    scale_s, scale_dQ, scale_dK, scale_dV, scale_dP = batched_args\n  query_bdim = batch_dims[0]\n  out_bdims = query_bdim, query_bdim, query_bdim\n\n  if layout == AttentionLayout.BNTH.value:\n    *Bs, N, T, _ = query.shape\n    *_, _, S, _ = key.shape\n  else:\n    *Bs, T, N, _ = query.shape\n    *_, S, _, _ = key.shape\n  B = math.prod(Bs)\n\n  # reshape to 4D shape\n  query = jnp.reshape(query, (B,) + query.shape[-3:])\n  key = jnp.reshape(key, (B,) + key.shape[-3:])\n  value = jnp.reshape(value, (B,) + key.shape[-3:])\n\n  activation = jnp.reshape(activation, (B, N, T))\n  fwd_output = jnp.reshape(fwd_output, (B,) + query.shape[-3:])\n  grad_output = jnp.reshape(grad_output, (B,) + query.shape[-3:])\n\n  grads = _dot_product_attention_fp8_bwd_p_wrapper.bind(\n      query, key, value, fwd_output, grad_output, activation,\n      descale_q, descale_k, descale_v, descale_o, descale_dO, descale_s, descale_dP, scale_s, scale_dQ, scale_dK, scale_dV, scale_dP,\n      scale=scale, use_causal_mask=use_causal_mask, layout=layout,\n  )\n\n  grad_query, grad_key, grad_value = grads[:3]\n  # reshape to original shape\n  grad_query = jnp.reshape(grad_query, query.shape)\n  grad_key = jnp.reshape(grad_key, key.shape)\n  grad_value = jnp.reshape(grad_value, value.shape)\n\n  return grads, out_bdims\n\ndef _infer_fp8_fwd_output_sharding(mesh, arg_shapes, is_training, layout):\n  # Prepare variadic_args for the original function\n  has_bias = False  # Adjust as needed\n  variadic_args = (has_bias, None)  # Dummy value, adjust as necessary\n\n  # Call the original function with the required parameters\n  output_sharding = _infer_fwd_output_sharding(mesh, arg_shapes, variadic_args, is_training, layout)\n  amax_sharding = NamedSharding(mesh, PartitionSpec())\n  if is_training:\n    out_sharding, activation_sharding = output_sharding[0], output_sharding[1]\n    return [out_sharding, amax_sharding, amax_sharding, activation_sharding]\n  return output_sharding + [amax_sharding, amax_sharding]\n\n_dot_product_attention_fp8_fwd_lower = custom_partitioning(\n    _dot_product_attention_fp8_fwd_impl, static_argnums=(9, 10, 11, 12))\n\ndef _dot_product_attention_fp8_fwd_infer_sharding_from_operands(\n    scale, use_causal_mask, layout, is_training,\n    mesh, arg_shapes, result_shape):\n  return _infer_fp8_fwd_output_sharding(mesh, arg_shapes, is_training, layout)\n\ndef _dot_product_attention_fp8_fwd_partition(\n    scale, use_causal_mask, layout, is_training,\n    mesh, arg_shapes, result_shape):\n  # args sharding\n  arg_shardings = tuple(arg_i.sharding for arg_i in arg_shapes)\n  out_shardings = _infer_fp8_fwd_output_sharding(\n    mesh, arg_shapes, is_training, layout)\n  impl = functools.partial(\n      _dot_product_attention_fp8_fwd_impl, scale=scale, use_causal_mask=use_causal_mask,\n      layout=layout, is_training=is_training)\n  return mesh, impl, out_shardings, arg_shardings\n\ndef _infer_fp8_bwd_output_sharding(mesh, arg_shapes, layout):\n  # Prepare variadic_args for the original function\n  has_bias = False  # Adjust as needed\n  has_dbias = False  # Adjust as needed\n  variadic_args = (has_bias, has_dbias)  # Dummy value, adjust as necessary\n\n  # Call the original function with the required parameters\n  output_shardings = _infer_bwd_output_sharding(mesh, arg_shapes, layout, variadic_args)\n\n  # Prepare amax_sharding\n  amax_sharding = NamedSharding(mesh, PartitionSpec())  # Use a default spec or adjust as needed\n\n  # Append amax_sharding for each output sharding\n  out_shardings_with_amax = output_shardings + [amax_sharding] * 4\n\n  return out_shardings_with_amax\n\n_dot_product_attention_fp8_bwd_lower = custom_partitioning(\n    _dot_product_attention_fp8_bwd_impl, static_argnums=(18,19,20)\n)\n\ndef _dot_product_attention_fp8_bwd_infer_sharding_from_operands(\n    scale, use_causal_mask, layout, mesh,\n    arg_shapes, result_shape):\n  return _infer_fp8_bwd_output_sharding(mesh, arg_shapes, layout)\n\ndef _dot_product_attention_fp8_bwd_partition(\n    scale, use_causal_mask, layout, mesh,\n    arg_shapes, result_shape):\n  out_shardings = _infer_fp8_bwd_output_sharding(mesh, arg_shapes, layout)\n  # args sharding\n  arg_shardings = tuple(arg_i.sharding for arg_i in arg_shapes)\n  impl = functools.partial(\n      _dot_product_attention_fp8_bwd_impl, scale=scale,\n      use_causal_mask=use_causal_mask, layout=layout\n  )\n  return mesh, impl, out_shardings, arg_shardings\n\n# Create dot_product_attention_fp8_fwd_p for forward operation.\n_dot_product_attention_fp8_fwd_p = core.Primitive(""dot_product_attention_fp8_fwd"")\n_dot_product_attention_fp8_fwd_p.multiple_results = True\n_dot_product_attention_fp8_fwd_p.def_impl(\n    functools.partial(dispatch.apply_primitive, _dot_product_attention_fp8_fwd_p)\n)\n_dot_product_attention_fp8_fwd_p.def_abstract_eval(\n    _dot_product_attention_fp8_fwd_abstract\n)\n\nmlir.register_lowering(\n  _dot_product_attention_fp8_fwd_p,\n  _dot_product_attention_fp8_fwd_cuda_lowering,\n  platform=""cuda"",\n)\n\n_dot_product_attention_fp8_fwd_p_wrapper = core.Primitive(\n    ""dot_product_attention_fp8_fwd_wrapper""\n)\n_dot_product_attention_fp8_fwd_p_wrapper.multiple_results = True\n_dot_product_attention_fp8_fwd_p_wrapper.def_impl(_dot_product_attention_fp8_fwd_impl)\n_dot_product_attention_fp8_fwd_p_wrapper.def_abstract_eval(\n    _dot_product_attention_fp8_fwd_abstract\n)\n\n# Create dot_product_attention_bwd_p for backward operation.\n_dot_product_attention_fp8_bwd_p = core.Primitive(""dot_product_attention_fp8_bwd"")\n_dot_product_attention_fp8_bwd_p.multiple_results = True\n_dot_product_attention_fp8_bwd_p.def_impl(\n    functools.partial(dispatch.apply_primitive, _dot_product_attention_fp8_bwd_p)\n)\n_dot_product_attention_fp8_bwd_p.def_abstract_eval(\n    _dot_product_attention_fp8_bwd_abstract\n)\n\nmlir.register_lowering(\n  _dot_product_attention_fp8_bwd_p,\n  _dot_product_attention_fp8_bwd_cuda_lowering,\n  platform=""cuda"",\n)\n\n_dot_product_attention_fp8_bwd_p_wrapper = core.Primitive(\n    ""dot_product_attention_fp8_bwd_wrapper""\n)\n_dot_product_attention_fp8_bwd_p_wrapper.multiple_results = True\n_dot_product_attention_fp8_bwd_p_wrapper.def_impl(_dot_product_attention_fp8_bwd_impl)\n_dot_product_attention_fp8_bwd_p_wrapper.def_abstract_eval(\n    _dot_product_attention_fp8_bwd_abstract\n)\n\nbatching.primitive_batchers[\n    _dot_product_attention_fp8_fwd_p_wrapper\n] = _dot_product_attention_fp8_fwd_batcher\nbatching.primitive_batchers[\n    _dot_product_attention_fp8_bwd_p_wrapper\n] = _dot_product_attention_fp8_bwd_batcher\n\n_dot_product_attention_fp8_fwd_lower.def_partition(\n  infer_sharding_from_operands=_dot_product_attention_fp8_fwd_infer_sharding_from_operands,\n  partition=_dot_product_attention_fp8_fwd_partition)\n\nmlir.register_lowering(_dot_product_attention_fp8_fwd_p_wrapper,\n                        mlir.lower_fun(_dot_product_attention_fp8_fwd_lower, multiple_results=True))\n\n_dot_product_attention_fp8_bwd_lower.def_partition(\n  infer_sharding_from_operands=_dot_product_attention_fp8_bwd_infer_sharding_from_operands,\n  partition=_dot_product_attention_fp8_bwd_partition)\n\nmlir.register_lowering(_dot_product_attention_fp8_bwd_p_wrapper,\n                        mlir.lower_fun(_dot_product_attention_fp8_bwd_lower, multiple_results=True))\n\ndispatch.prim_requires_devices_during_lowering.add(\n  _dot_product_attention_fp8_fwd_p\n)\ndispatch.prim_requires_devices_during_lowering.add(\n  _dot_product_attention_fp8_fwd_p_wrapper\n)\ndispatch.prim_requires_devices_during_lowering.add(\n  _dot_product_attention_fp8_bwd_p\n)\ndispatch.prim_requires_devices_during_lowering.add(\n  _dot_product_attention_fp8_bwd_p_wrapper\n)\n\n@functools.partial(jax.custom_vjp, nondiff_argnums=(4, 5, 6, 7))\ndef _dot_product_attention_fp8(query: Array,\n                               key: Array,\n                               value: Array,\n                               fp8_params: dict[str, Array],\n                               scale: float,\n                               use_causal_mask: bool,\n                               layout: int,\n                               cudnn_version: int):\n  output, amax_s, amax_o = _dot_product_attention_fp8_fwd(\n      query, key, value, params_from_keys(fp8_params, fp8_params_keys_fwd),\n      scale, use_causal_mask, layout, cudnn_version\n  )\n  return output, amax_s, amax_o\n\n_dot_product_attention_fp8.defvjp(_dot_product_attention_fp8_fwd_rule, _dot_product_attention_fp8_bwd_rule)\n\ndef combine_bias_and_mask(bias, mask, dtype):\n  if bias is not None:\n    # reshape bias to have 4D shape\n    bias = bias.reshape((1,) * (4 - len(bias.shape)) + bias.shape)\n\n  if mask is not None:\n    if mask.dtype == jnp.bool:\n      large_negative_number = get_large_negative_number(dtype)\n      mask = jnp.where(mask, jnp.asarray(0, dtype), large_negative_number)\n    # reshape mask to have 4D shape\n    mask = mask.reshape((1,) * (4 - len(mask.shape)) + mask.shape)  # type: ignore[union-attr]\n\n  # combine bias and mask\n  if bias is None:\n    bias = mask\n  else:\n    if mask is not None:\n      # should be broadcast to same shape\n      bias = bias + mask\n  return bias\n\n# User interface\ndef paged_attention(\n    query: Array,\n    key: Array,\n    value: Array,\n    q_seqlen: Array,\n    kv_seqlen: Array,\n    page_table_k: Array,\n    page_table_v: Array,\n    bias: Array | None = None,\n    mask: Array | None = None,\n    fp8_params: FP8Params | None = None,\n    *,\n    scale: float = 1.0,\n    mask_type: MaskType = MaskType.NO_MASK,\n    seed: int = 42,\n    dropout_rate: float = 0.,\n    qkv_layout: str = ""BTNH"",\n    sliding_window_length: int | None = None,\n    use_fp8: bool = False,\n    return_residual: bool = False\n):\n  """"""Computes paged attention described in https://arxiv.org/pdf/2309.06180.\n\n    B = batch size\n    S = length of the key/value (source)\n    T = length of the query (target)\n    N = number of attention heads\n    H = dimensions of each attention head.\n\n  Args:\n    query: Queries for attention calculation with a shape of BTNH or BNTH.\n    key: Keys for attention calculation with a shape of\n      [num_blocks, block_size, N, H] or [num_blocks, N, block_size, H] where\n      num_blocks = B * Ceil(S / block_size).\n    value: Values to be used in attention with a shape of\n      [num_blocks, block_size, N, H] or [num_blocks, N, block_size, H] where\n      num_blocks = B * Ceil(S / block_size).\n    q_seqlen: Non padded sequence length of query with a shape of B.\n    kv_seqlen: Non padded sequence length of key and value with a shape of B.\n    page_table_k: page table for key of shape [B, 1, num_blocks_per_batch, 1]\n      where num_blocks_per_batch = Ceil(S / block_size).\n    page_table_v: page table for value of shape [B, 1, num_blocks_per_batch, 1]\n      where num_blocks_per_batch = Ceil(S / block_size).\n    bias: Bias to be added to logits with a shape of BNTS.\n    mask: Mask used to filter out logits with a shape of BNTS.\n    scale: Scale for the query.\n    qkv_layout: Layout string, with supported formats being BTNH, BNTH, BSNH,\n      BNSH.\n    sliding_window_length: Window size to make attention only attend to each\n      token's left local window (pos - sliding_window_length, pos] where `pos`\n      is the index of each token. E.g., if sliding_window_length == 3 and the\n      sequence is [0, 1, 2, 3, c, 4, 5], token `c` can attend to [4, 5, c].\n    use_fp8: Whether to use FP8 attention mechanism.\n    return_residual: Whether to return the logsumexp tensor of shape BTN\n      or BNT to users. See section 3.1.1 in the FlashAttention-2 paper:\n      https://arxiv.org/pdf/2307.08691 to find the definition of logsumexp.\n  Returns:\n    output: the same shape as the query.\n    residual: the logsumexp tensor if return_residual=True. (non fp8)\n  """"""\n  cudnn_version = check_cudnn_version()\n  layout = _normalize_layout(qkv_layout)\n  if use_fp8:\n    raise ValueError(""Paged attention doesn't support fp8 for now."")\n  if has_padding(mask_type) and (q_seqlen is None or kv_seqlen is None):\n    raise ValueError(""Require q_seqlen and kv_seqlen to generate padding mask."")\n  if sliding_window_length is not None and sliding_window_length <= 0:\n    raise ValueError(\n      f""Require sliding_window_length > 0, got {sliding_window_length}."")\n\n  bias = combine_bias_and_mask(bias, mask, query.dtype)\n  # check if input shape and data type is compatiable\n  check_layout(query, key, value, bias, q_seqlen, kv_seqlen, None, None,\n    page_table_k, page_table_v, layout)\n  has_bias = bias is not None\n  has_dbias = has_bias and \\n    should_export_dbias(bias.shape, query.shape, layout)  # type: ignore[union-attr]\n  variadic_args = (has_bias, has_dbias)\n\n  _not_used = jnp.zeros(0, dtype=query.dtype)\n  if bias is None:\n    bias = _not_used\n\n  output = _dot_product_attention(\n      query, key, value, bias, q_seqlen, kv_seqlen, _not_used, _not_used,\n      page_table_k, page_table_v, scale, seed, dropout_rate, variadic_args,\n      mask_type, layout.value, sliding_window_length, cudnn_version,\n      return_residual)\n  return output\n\n\ndef dot_product_attention(\n    query: Array,\n    key: Array,\n    value: Array,\n    bias: Array | None = None,\n    mask: Array | None = None,\n    q_seqlen: Array | None = None,\n    kv_seqlen: Array | None = None,\n    q_offsets: Array | None = None,\n    kv_offsets: Array | None = None,\n    fp8_params: FP8Params | None = None,\n    *,\n    scale: float = 1.0,\n    mask_type: MaskType = MaskType.NO_MASK,\n    seed: int = 42,\n    dropout_rate: float = 0.,\n    qkv_layout: str = ""BTNH"",\n    sliding_window_length: int | None = None,\n    use_fp8: bool = False,\n    return_residual: bool = False\n):\n  """"""Computes dot-product attention given query (Q), key (K), and value (V).\n\n  This function serves as the core operation for applying attention\n  mechanisms as described in the paper [https://arxiv.org/abs/1706.03762].\n  Initially, it determines the attention weights by processing Q and K,\n  subsequently combining the outcomes using K. Throughout this function, we\n  utilize the following uppercase letters to represent specific parameters of\n  array:\n\n    B = batch size\n    S = length of the key/value (source)\n    T = length of the query (target)\n    N = number of attention heads\n    H = dimensions of each attention head.\n\n  The supported layouts for Q, K, V are either BT(S)NH or BNT(S)H, and they must\n  adhere to the same layout. The output layout remains consistent with Q,\n  defaulting to BT(S)NH.\n\n  Args:\n    query: Queries for attention calculation with a shape of BTNH or BNTH.\n    key: Keys for attention calculation with a shape of BSNH or BNSH.\n    value: Values to be used in attention with a shape of BSNH or BNSH.\n    bias: Bias to be added to logits with a shape of BNTS.\n    mask: Mask used to filter out logits with a shape of BNTS.\n    q_seqlen: Non padded sequence length of query with a shape of B.\n      If q_offsets is set, q_seqlen should have shape [B,M] where M is the\n      maximum number of segments per batch. For batch that has less segments\n      than maximum segments, fill the padded entries with -1.\n    kv_seqlen: Non padded sequence length of key and value with a shape of B.\n      If kv_offsets is set, kv_seqlen should have shape [B,M] where M is the\n      maximum number of segments per batch. For batch that has less segments\n      than maximum segments, fill the padded entries with -1.\n    q_offsets: offset of each segment packed in query with a shape of [B,M+1]\n      where M is the maximum number of segments per batch. For batch that has\n      less segments than maximum segments, fill the padded entries with -1.\n      E.g, if 2 batches has 3 and 2 segments respectively, each segment has\n      size 1, q_offsets = [[0,1,2,-1], [0,1,-1,-1]]. q_seqlen should be set\n      to indicate the size of each segment.\n    kv_offsets: offset of each segment packed in key with a shape of [B,M+1]\n      where M is the maximum number of segments per batch. For batch that has\n      less segments than maximum segments, fill the padded entries with -1.\n      E.g, if 2 batches has 3 and 2 segments respectively, each segment has\n      size 1, kv_offsets = [[0,1,2,-1], [0,1,-1,-1]]. kv_seqlen should be set\n      to indicate the size of each segment.\n    scale: Scale for the query.\n    dropout_rate: Dropout rate.\n    qkv_layout: Layout string, with supported formats being BTNH, BNTH, BSNH,\n      BNSH.\n    sliding_window_length: Window size to make attention only attend to each\n      token's left local window (pos - sliding_window_length, pos] where `pos`\n      is the index of each token. E.g., if sliding_window_length == 3 and the\n      sequence is [0, 1, 2, 3, c, 4, 5], token `c` can attend to [4, 5, c].\n    use_fp8: Whether to use FP8 attention mechanism.\n    return_residual: Whether to return the logsumexp tensor of shape BTN\n      or BNT to users. See section 3.1.1 in the FlashAttention-2 paper:\n      https://arxiv.org/pdf/2307.08691 to find the definition of logsumexp.\n  Returns:\n    output: the same shape as the query.\n    residual: the logsumexp tensor if return_residual=True. (non fp8)\n    amax_s: amax of state. (fp8 only)\n    amax_o: amax of output. (fp8 only)\n  """"""\n  # TODO(b/380898464): Check the compute capability, e.g., require GPU device,\n  # in the kernel implementation (c++) code.\n  cudnn_version = check_cudnn_version()\n  layout = _normalize_layout(qkv_layout)\n\n  if use_fp8:\n    if fp8_params is None:\n      raise ValueError(""fp8_params should not be None."")\n    if  mask_type not in (MaskType.NO_MASK, MaskType.CAUSAL):\n      raise ValueError(""Only NO_MASK or CAUSAL masks are supported for fp8."")\n    if not all(x is None for x in [bias, mask, q_seqlen, kv_seqlen]):\n      raise ValueError(\n          f""Expected 'None' for bias, mask, q_seqlen, and kv_seqlen, ""\n          f""but got: bias={bias}, mask={mask}, q_seqlen={q_seqlen}, kv_seqlen={kv_seqlen}""\n      )\n    check_fp8_params(fp8_params)\n    check_layout(query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n      None, None, layout)\n    output, amax_s, amax_o = _dot_product_attention_fp8(\n        query, key, value, fp8_params,\n        scale, mask_type == MaskType.CAUSAL, layout.value, cudnn_version\n    )\n    return output, amax_s, amax_o\n  else:\n    if has_padding(mask_type) and (q_seqlen is None or kv_seqlen is None):\n        raise ValueError(""Require q_seqlen and kv_seqlen to generate padding mask"")\n    if sliding_window_length is not None and sliding_window_length <= 0:\n      raise ValueError(\n        f""Require sliding_window_length > 0, got {sliding_window_length}"")\n    if q_offsets is not None and (q_seqlen is None or kv_seqlen is None):\n      raise ValueError(""Require q_seqlen and kv_seqlen to use packed layout"")\n\n    bias = combine_bias_and_mask(bias, mask, query.dtype)\n    # check if input shape and data type is compatiable\n    check_layout(query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n      None, None, layout)\n    has_bias = bias is not None\n    has_dbias = has_bias and \\n      should_export_dbias(bias.shape, query.shape, layout)  # type: ignore[union-attr]\n    variadic_args = (has_bias, has_dbias)\n\n    _not_used = jnp.zeros(0, dtype=query.dtype)\n    if bias is None:\n      bias = _not_used\n    if q_seqlen is None:\n      q_seqlen = _not_used\n    if kv_seqlen is None:\n      kv_seqlen = _not_used\n    if q_offsets is None:\n      q_offsets = _not_used\n    if kv_offsets is None:\n      kv_offsets = _not_used\n\n    output = _dot_product_attention(\n        query, key, value, bias, q_seqlen, kv_seqlen, q_offsets, kv_offsets,\n        _not_used, _not_used, scale, seed, dropout_rate, variadic_args,\n        mask_type, layout.value, sliding_window_length, cudnn_version,\n        return_residual)\n    return output\n",python,tab
+2,337,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"6:59:28 AM [info] Activating crowd-code\n6:59:28 AM [info] Recording started\n6:59:28 AM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,637,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"6:59:28 AM [info] Git repository found\n6:59:28 AM [info] Git provider initialized successfully\n6:59:29 AM [info] Initial git state: [object Object]\n",Log,content
+4,175136,".venv/lib/python3.10/site-packages/jax/_src/cudnn/fused_attention_stablehlo.py",0,0,"",python,tab
+5,237672,".venv/lib/python3.10/site-packages/jax/_src/cudnn/fused_attention_stablehlo.py",12124,0,"",python,selection_command
+6,238004,".venv/lib/python3.10/site-packages/jax/_src/cudnn/fused_attention_stablehlo.py",12107,19,"",python,content
+7,238080,".venv/lib/python3.10/site-packages/jax/_src/cudnn/fused_attention_stablehlo.py",12113,0,"",python,selection_command
+8,239005,".venv/lib/python3.10/site-packages/jax/_src/cudnn/fused_attention_stablehlo.py",12084,0,"",python,selection_command
+9,239210,".venv/lib/python3.10/site-packages/jax/_src/cudnn/fused_attention_stablehlo.py",12054,0,"",python,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-209587dd-d381-4d16-adb9-65d8ac8ce07c1755718881164-2025_08_20-21.41.26.905/source.csv

@@ -0,0 +1,4 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"test.py",0,0,"import jax\nimport jax.numpy as jnp\nimport numpy as np\n\njax.config.update(""jax_transfer_guard"", ""disallow"")\n\nx_on_cpu = jax.device_put(1)\n\nprint(float(x_on_cpu))",python,tab
+2,123,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"9:41:26 PM [info] Activating crowd-code\n9:41:26 PM [info] Recording started\n9:41:26 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,158,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"9:41:26 PM [info] Git repository found\n9:41:26 PM [info] Git provider initialized successfully\n9:41:27 PM [info] Initial git state: [object Object]\n",Log,content

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-38ccda13-4be3-4e62-821a-55fb07445dc01765228671121-2025_12_08-22.17.58.352/source.csv

@@ -0,0 +1,109 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,422,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"10:17:58 PM [info] Activating crowd-code\n10:17:58 PM [info] Recording started\n10:17:58 PM [info] Initializing git provider using file system watchers...\n10:17:58 PM [info] No workspace folder found\n",Log,tab
+3,2292,"extension-output-pdoom-org.crowd-code-#1-crowd-code",198,0,"10:18:00 PM [info] Retrying git provider initialization...\n10:18:00 PM [info] No workspace folder found\n",Log,content
+4,10561,"Untitled-1",0,0,"",plaintext,tab
+5,13510,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",0,0,"#!/usr/bin/env python3\nimport sys\nfrom pathlib import Path\nimport pandas as pd\n\ninput_dir = sys.argv[1]\nfor parquet_file in sorted(Path(input_dir).glob(""shard_*.parquet"")):\n    df = pd.read_parquet(parquet_file)\n    for text in df['text']:\n        print(text)\n",python,tab
+6,18949,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",260,0,"\n",python,content
+7,22077,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",261,0,"i",python,content
+8,22079,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",262,0,"",python,selection_keyboard
+9,22272,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",262,0,"f",python,content
+10,22275,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",263,0,"",python,selection_keyboard
+11,22324,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",263,0," ",python,content
+12,22327,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",264,0,"",python,selection_keyboard
+13,43753,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",264,0,"_",python,content
+14,43755,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",265,0,"",python,selection_keyboard
+15,44863,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",265,0,"_",python,content
+16,44866,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",266,0,"",python,selection_keyboard
+17,45269,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",266,0,"n",python,content
+18,45272,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",267,0,"",python,selection_keyboard
+19,45438,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",267,0,"a",python,content
+20,45441,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",268,0,"",python,selection_keyboard
+21,45485,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",268,0,"m",python,content
+22,45487,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",269,0,"",python,selection_keyboard
+23,45566,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",269,0,"e",python,content
+24,45569,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",270,0,"",python,selection_keyboard
+25,45940,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",270,0,"_",python,content
+26,45943,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",271,0,"",python,selection_keyboard
+27,46118,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",271,0,"_",python,content
+28,46121,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",272,0,"",python,selection_keyboard
+29,49306,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",272,0," ",python,content
+30,49308,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",273,0,"",python,selection_keyboard
+31,55687,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",273,0,"=",python,content
+32,55690,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",274,0,"",python,selection_keyboard
+33,55944,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",274,0," ",python,content
+34,55947,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",275,0,"",python,selection_keyboard
+35,72432,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",275,0,"""""",python,content
+36,72435,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",276,0,"",python,selection_keyboard
+37,72570,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",276,1,"""",python,content
+38,72572,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",277,0,"",python,selection_keyboard
+39,73101,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",276,0,"",python,selection_command
+40,77820,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",261,16,"if __name__ == ""__main__"":\n",python,content
+41,84743,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",288,0,"    print(""Running as main"")\n",python,content
+42,101748,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",317,0,"    print(f""Processing input directory: {input_dir}"")\n",python,content
+43,124236,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",104,157,"",python,content
+44,137889,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",214,0,"for parquet_file in sorted(Path(input_dir).glob(""shard_*.parquet"")):\n    df = pd.read_parquet(parquet_file)\n    for text in df[\'text\']:\n        print(text)\n",python,content
+45,146797,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",372,0,"    for parquet_file in sorted(Path(input_dir).glob(""shard_*.parquet"")):\n        df = pd.read_parquet(parquet_file)\n        for text in df[\'text\']:\n            print(text)\n",python,content
+46,152488,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",146,0,"",python,selection_command
+47,152630,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",175,0,"",python,selection_command
+48,152770,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",229,0,"",python,selection_command
+49,159053,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",228,0,"",python,selection_command
+50,159589,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",214,68,"for parquet_file in sorted(Path(input_dir).glob(""shard_*.parquet"")):",python,selection_command
+51,159784,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",214,107,"for parquet_file in sorted(Path(input_dir).glob(""shard_*.parquet"")):\n    df = pd.read_parquet(parquet_file)",python,selection_command
+52,159921,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",214,137,"for parquet_file in sorted(Path(input_dir).glob(""shard_*.parquet"")):\n    df = pd.read_parquet(parquet_file)\n    for text in df[\'text\']:",python,selection_command
+53,160055,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",214,157,"for parquet_file in sorted(Path(input_dir).glob(""shard_*.parquet"")):\n    df = pd.read_parquet(parquet_file)\n    for text in df[\'text\']:\n        print(text)",python,selection_command
+54,168793,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",214,158,"",python,content
+55,168799,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",218,0,"",python,selection_command
+56,171710,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",291,0,"",python,selection_command
+57,172026,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",334,0,"",python,selection_command
+58,172029,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",368,0,"",python,selection_command
+59,172032,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",388,0,"",python,selection_command
+60,179863,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",368,0,"",python,selection_command
+61,180097,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",334,0,"",python,selection_command
+62,180134,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",291,0,"",python,selection_command
+63,180164,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",218,0,"",python,selection_command
+64,180198,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",164,0,"",python,selection_command
+65,180706,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",135,0,"",python,selection_command
+66,187273,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",131,28,"    print(""Running as main"")",python,selection_command
+67,187549,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",131,82,"    print(""Running as main"")\n    print(f""Processing input directory: {input_dir}"")",python,selection_command
+68,188659,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",131,83,"",python,content
+69,188665,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",135,0,"",python,selection_command
+70,192944,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",108,0,"",python,selection_command
+71,193107,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",84,0,"",python,selection_command
+72,193627,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",80,24,"",python,content
+73,199284,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",107,0,"    input_dir = sys.argv[1]\n",python,content
+74,214257,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",111,28,"",python,content
+75,214503,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",81,0,"nput_dir = sys.argv[1]\ni",python,content
+76,214506,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",84,0,"",python,selection_command
+77,214550,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",135,0,"print(""Running as main"")\n    print(f""Processing input directory: {input_dir}"")\n    ",python,content
+78,214554,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",135,0,"",python,selection_command
+79,214566,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",214,0,"for parquet_file in sorted(Path(input_dir).glob(""shard_*.parquet"")):\n    df = pd.read_parquet(parquet_file)\n    for text in df[\'text\']:\n        print(text)\n",python,content
+80,214570,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",228,0,"",python,selection_command
+81,214602,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",372,174,"\nif __name__ = """"",python,content
+82,214606,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",347,1,"",python,content
+83,214610,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",341,1,"",python,content
+84,214613,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",104,110,"",python,content
+85,214619,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",119,0,"",python,selection_command
+86,214637,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",260,17,"",python,content
+87,214645,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",260,0,"",python,selection_command
+88,219798,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",260,0,"\nif __name__ = """"",python,content
+89,219805,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",260,0,"",python,selection_command
+90,220049,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",104,0,"if __name__ == ""__main__"":\n    print(""Running as main"")\n    print(f""Processing input directory: {input_dir}"")\n",python,content
+91,220055,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",341,0,"\",python,content
+92,220059,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",347,0,"\",python,content
+93,220061,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",372,17,"    for parquet_file in sorted(Path(input_dir).glob(""shard_*.parquet"")):\n        df = pd.read_parquet(parquet_file)\n        for text in df[\'text\']:\n            print(text)\n",python,content
+94,220064,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",119,0,"",python,selection_command
+95,220069,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",214,158,"",python,content
+96,220071,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",228,0,"",python,selection_command
+97,220099,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",135,83,"",python,content
+98,220104,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",135,0,"",python,selection_command
+99,220131,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",81,24,"",python,content
+100,220137,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",84,0,"",python,selection_command
+101,220165,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",111,0,"input_dir = sys.argv[1]\n    ",python,content
+102,223940,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",111,0,"",python,selection_command
+103,224191,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",139,0,"",python,selection_command
+104,224224,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",212,0,"",python,selection_command
+105,224384,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",255,0,"",python,selection_command
+106,224564,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",289,0,"",python,selection_command
+107,224732,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",255,0,"",python,selection_command
+108,227689,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",251,33,"        for text in df[\'text\']:",python,selection_command
+109,228194,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",255,0,"",python,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-3acc90e9-90ce-4c91-8dc5-7fa36ee6eae81754056616784-2025_08_01-15.57.02.654/source.csv

@@ -0,0 +1,8 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"experiments/sample.sh",0,0,"source .venv/bin/activate\n\ndata_dir=""$PWD/data_arrayrecord/dummy""\nckpt_dir=""$PWD/checkpoints/causal_dynamics_openai_grain_tok_restore""\n\nexport PYTHONUNBUFFERED=1\nsrun ipython --pdb sample.py -- \\n    --dyna_type ""causal"" \\n    --batch_size 1 \\n    --seq_len 2 \\n    --start_frame 1 \\n    --checkpoint $ckpt_dir \\n    --data_dir $data_dir",shellscript,tab
+2,589,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"3:57:02 PM [info] Activating crowd-code\n3:57:02 PM [info] Recording started\n3:57:02 PM [info] Initializing git provider using file system watchers...\n3:57:02 PM [info] Git repository found\n3:57:02 PM [info] Git provider initialized successfully\n3:57:02 PM [info] Initial git state: [object Object]\n",Log,tab
+3,750,"TERMINAL",0,0,"",,terminal_focus
+4,910,"experiments/sample.sh",0,0,"",shellscript,tab
+5,8633,"utils/nn.py",0,0,"import math\nfrom typing import Tuple, Callable, List\n\nfrom flax import nnx\nimport jax\nimport jax.numpy as jnp\nimport einops\n\n\nclass PositionalEncoding(nnx.Module):\n    """"""https://uvadlc-notebooks.readthedocs.io/en/latest/tutorial_notebooks/JAX/tutorial6/Transformers_and_MHAttention.html""""""\n\n    def __init__(self, d_model: int, max_len: int = 5000):\n        self.d_model = d_model\n        self.max_len = max_len\n\n        pe = jnp.zeros((self.max_len, self.d_model))\n        position = jnp.arange(0, self.max_len, dtype=jnp.float32)[:, None]\n        div_term = jnp.exp(\n            jnp.arange(0, self.d_model, 2) * (-math.log(10000.0) / self.d_model)\n        )\n        pe = pe.at[:, 0::2].set(jnp.sin(position * div_term))\n        pe = pe.at[:, 1::2].set(jnp.cos(position * div_term))\n        self.pe = nnx.Variable(pe)\n\n    def __call__(self, x: jax.Array) -> jax.Array:\n        x = x + self.pe[: x.shape[2]]\n        return x\n\n\nclass STBlock(nnx.Module):\n    def __init__(\n        self,\n        dim: int,\n        ffn_dim: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        rngs: nnx.Rngs,\n    ):\n        self.dim = dim\n        self.ffn_dim = ffn_dim\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n\n        self.spatial_pos_enc = PositionalEncoding(self.dim)\n        self.spatial_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.spatial_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.dim,\n            qkv_features=self.dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=False\n            ),\n            rngs=rngs,\n            decode=False,\n        )\n\n        self.temporal_pos_enc = PositionalEncoding(self.dim)\n        self.temporal_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.temporal_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.dim,\n            qkv_features=self.dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=False,\n        )\n\n        self.ffn_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense1 = nnx.Linear(\n            in_features=self.dim,\n            out_features=self.ffn_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense2 = nnx.Linear(\n            in_features=self.ffn_dim,\n            out_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    @nnx.remat\n    def __call__(self, x_BTNM: jax.Array) -> jax.Array:\n        # --- Spatial attention ---\n        z_BTNM = self.spatial_pos_enc(x_BTNM)\n        z_BTNM = self.spatial_norm(z_BTNM)\n        z_BTNM = self.spatial_attention(z_BTNM)\n        x_BTNM = x_BTNM + z_BTNM\n\n        # --- Temporal attention ---\n        x_BNTM = x_BTNM.swapaxes(1, 2)\n        z_BNTM = self.temporal_pos_enc(x_BNTM)\n        z_BNTM = self.temporal_norm(z_BNTM)\n        z_BNTM = self.temporal_attention(z_BNTM)\n        x_BNTM = x_BNTM + z_BNTM\n        x_BTNM = x_BNTM.swapaxes(1, 2)\n\n        # --- Feedforward ---\n        z_BTNM = self.ffn_norm(x_BTNM)\n        z_BTND = self.ffn_dense1(z_BTNM)\n        z_BTND = jax.nn.gelu(z_BTND)\n        z_BTNM = self.ffn_dense2(z_BTND)\n        x_BTNM = x_BTNM + z_BTNM\n\n        return x_BTNM\n\n\nclass STTransformer(nnx.Module):\n    """"""\n    Dimension keys:\n        B: batch size\n        T: number of frames\n        N: number of patches per frame\n        I: number of input features\n        M: model dimension\n        D: FFN dimension\n        O: number of output features\n    """"""\n    def __init__(\n        self,\n        input_dim: int,\n        model_dim: int,\n        ffn_dim: int,\n        out_dim: int,\n        num_blocks: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        rngs: nnx.Rngs,\n    ):\n        self.input_dim = input_dim\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.out_dim = out_dim\n        self.num_blocks = num_blocks\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n\n        self.input_norm1 = nnx.LayerNorm(\n            num_features=self.input_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_dense = nnx.Linear(\n            in_features=self.input_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_norm2 = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n        self.blocks = []\n        for _ in range(self.num_blocks):\n            self.blocks.append(\n                STBlock(\n                    dim=self.model_dim,\n                    ffn_dim=self.ffn_dim,\n                    num_heads=self.num_heads,\n                    dropout=self.dropout,\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                    use_flash_attention=self.use_flash_attention,\n                    rngs=rngs,\n                )\n            )\n\n        self.output_dense = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.out_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    def __call__(self, x_BTNI: jax.Array) -> jax.Array:\n        x_BTNI = self.input_norm1(x_BTNI)\n        x_BTNM = self.input_dense(x_BTNI)\n        x_BTNM = self.input_norm2(x_BTNM)\n\n        for block in self.blocks:\n            x_BTNM = block(x_BTNM)\n\n        x_BTNO = self.output_dense(x_BTNM)\n        return x_BTNO\n\nclass TransformerBlock(nnx.Module):\n    def __init__(\n        self,\n        model_dim: int,\n        ffn_dim: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        decode: bool,\n        rngs: nnx.Rngs,\n    ):\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.decode = decode\n\n        self.temporal_pos_enc = PositionalEncoding(self.model_dim)\n        self.spatial_pos_enc = PositionalEncoding(self.model_dim)\n        self.temporal_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.spatial_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.temporal_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.model_dim,\n            qkv_features=self.model_dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=self.decode,\n        )\n        self.spatial_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.model_dim,\n            qkv_features=self.model_dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=self.decode,\n        )\n        self.ffn_dense1 = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.ffn_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense2 = nnx.Linear(\n            in_features=self.ffn_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    @nnx.remat\n    def __call__(self, x_BTNM: jax.Array, pos_index: Tuple[jax.Array, jax.Array] | None = None) -> jax.Array:\n        # --- Spatial attention ---\n        B, T, N, M = x_BTNM.shape\n        z_FNM = einops.rearrange(x_BTNM, ""b t n m -> (b t) n m"")\n        z_FNM = self.spatial_norm(z_FNM)\n        if self.decode:\n            assert pos_index is not None\n            z_FM = z_FNM[:, pos_index[1]]\n            z_F1M = jnp.reshape(z_FM, (B * T, 1, M))\n            z_F1M = self.spatial_attention(z_F1M)\n            z_FM = jnp.reshape(z_F1M, (B * T, M))\n            z_FNM = z_FNM.at[:, pos_index[1], :].set(z_FM)\n        else:\n            z_FNM = self.spatial_attention(z_FNM)\n        z_BTNM = einops.rearrange(z_FNM, ""(b t) n m -> b t n m"", t=T)\n        x_BTNM = x_BTNM + z_BTNM\n        # --- Temporal attention ---\n        z_PTM = einops.rearrange(x_BTNM, ""b t n m -> (b n) t m"")\n        z_PTM = self.temporal_norm(z_PTM)\n        if self.decode:\n            assert pos_index is not None\n            z_PM = z_PTM[:, pos_index[0]]\n            z_P1M = jnp.reshape(z_PM, (B * N, 1, M))\n            z_P1M = self.temporal_attention(z_P1M)\n            z_PM = jnp.reshape(z_P1M, (B * N, M))\n            z_PTM = z_PTM.at[:, pos_index[0], :].set(z_PM)\n        else:\n            z_PTM = self.temporal_attention(z_PTM)\n        z_BTNM = einops.rearrange(z_PTM, ""(b n) t m -> b t n m"", n=N)\n        x_BTNM = x_BTNM + z_BTNM\n        # --- Feedforward ---\n        z_BTNM = self.ffn_norm(x_BTNM)\n        z_BTND = self.ffn_dense1(z_BTNM)\n        z_BTND = jax.nn.gelu(z_BTND)\n        z_BTNM = self.ffn_dense2(z_BTND)\n        x_BTNM = x_BTNM + z_BTNM\n\n        return x_BTNM\n\nclass Transformer(nnx.Module):\n    """"""\n    Dimension keys:\n        B: batch size\n        T: number of frames\n        N: number of patches per frame\n        I: number of input features\n        M: model dimension\n        D: FFN dimension\n        O: number of output features\n        F: number of frames in batch\n        P: number of patch positions in batch\n    """"""\n    def __init__(\n        self,\n        input_dim: int,\n        model_dim: int,\n        ffn_dim: int,\n        out_dim: int,\n        num_blocks: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        decode: bool,\n        rngs: nnx.Rngs,\n    ):\n        self.input_dim = input_dim\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.out_dim = out_dim\n        self.num_blocks = num_blocks\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n\n        self.pos_enc = PositionalEncoding(self.model_dim)\n        self.input_norm1 = nnx.LayerNorm(\n            num_features=self.input_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_dense = nnx.Linear(\n            in_features=self.input_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_norm2 = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n        self.blocks: List[TransformerBlock] = []\n        for _ in range(self.num_blocks):\n            self.blocks.append(\n                TransformerBlock(\n                    model_dim=self.model_dim,\n                    ffn_dim=self.ffn_dim,\n                    num_heads=self.num_heads,\n                    dropout=self.dropout,\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                    use_flash_attention=self.use_flash_attention,\n                    decode=decode,\n                    rngs=rngs,\n                )\n            )\n        self.output_dense = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.out_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    def __call__(self, x_BTNI: jax.Array, pos_index: Tuple[jax.Array, jax.Array] | None = None) -> jax.Array:\n        x_BTNI = self.input_norm1(x_BTNI)\n        x_BTNM = self.input_dense(x_BTNI)\n        x_BTNM = self.input_norm2(x_BTNM)\n        x_BTNM = self.pos_enc(x_BTNM)\n\n        for block in self.blocks:\n            x_BTNM = block(x_BTNM, pos_index)\n\n        x_BTNV = self.output_dense(x_BTNM)\n        return x_BTNV\n\ndef normalize(x: jax.Array) -> jax.Array:\n    return x / (jnp.linalg.norm(x, ord=2, axis=-1, keepdims=True) + 1e-8)\n\n\nclass VectorQuantizer(nnx.Module):\n    """"""\n    Dimension keys:\n        D: B * T * N\n        K: number of latents\n        L: latent dimension\n    """"""\n    def __init__(\n        self, latent_dim: int, num_latents: int, dropout: float, rngs: nnx.Rngs\n    ):\n        self.latent_dim = latent_dim\n        self.num_latents = num_latents\n        self.dropout = dropout\n\n        self.codebook = nnx.Param(\n            normalize(\n                nnx.initializers.lecun_uniform()(\n                    rngs.params(), (self.num_latents, self.latent_dim)\n                )\n            )\n        )\n        self.drop = nnx.Dropout(self.dropout, rngs=rngs)\n\n    def __call__(\n        self, x_DL: jax.Array, training: bool\n    ) -> Tuple[jax.Array, jax.Array, jax.Array, jax.Array]:\n        # --- Compute distances ---\n        x_DL = normalize(x_DL)\n        normalized_codebook_KL = normalize(self.codebook.value)\n        distance_DK = -jnp.matmul(x_DL, normalized_codebook_KL.T)\n        if training:\n            distance_DK = self.drop(distance_DK)\n\n        # --- Get indices and embeddings ---\n        indices_D = jnp.argmin(distance_DK, axis=-1)\n        z_DL = self.codebook[indices_D]\n\n        # --- Straight through estimator ---\n        z_q_DL = x_DL + jax.lax.stop_gradient(z_DL - x_DL)\n        return z_q_DL, z_DL, x_DL, indices_D\n\n    def get_codes(self, indices_E: jax.Array) -> jax.Array:\n        return self.codebook[indices_E]\n\n\ndef _create_flash_attention_fn(use_flash_attention: bool, is_causal: bool) -> Callable:\n    """"""\n    Create an attention function that uses flash attention if enabled.\n\n    flax.nnx.MultiHeadAttention provides tensors with shape (batch..., length, num_heads, head_dim),\n    but jax.nn.dot_product_attention expects (batch, length, num_heads, head_dim). We reshape to\n    ensure compatibility. cuDNN's flash attention additionally requires a sequence length that\n    is a multiple of 4. We pad the sequence length to the nearest multiple of 4 and mask\n    accordingly. Note that cuDNN requires the mask to be broadcast before calling the attention\n    function due to strict shape checking.\n    """"""\n\n    # FIXME (f.srambical): keys and values could have different dimensionalities\n    def attention_fn(query_BSHD, key_BSHD, value_BSHD, bias=None, mask_B111=None, **kwargs):\n        implementation = ""cudnn"" if use_flash_attention else None\n\n        def _merge_batch_dims(x):\n            return einops.rearrange(x, ""... l h k -> (...) l h k"")\n\n        def _pad(x):\n            return jnp.pad(x, ((0, 0), (0, pad_size), (0, 0), (0, 0)))\n\n        original_shape = query_BSHD.shape\n        original_seq_len = query_BSHD.shape[-3]\n\n        # Pad to nearest multiple of 4\n        T = ((original_seq_len + 3) // 4) * 4\n        pad_size = T - original_seq_len\n\n        query_BTHD = _pad(_merge_batch_dims(query_BSHD))\n        key_BTHD = _pad(_merge_batch_dims(key_BSHD))\n        value_BTHD = _pad(_merge_batch_dims(value_BSHD))\n        B = query_BTHD.shape[0]\n\n        attention_mask = jnp.ones((T, T), dtype=jnp.bool_)\n        attention_mask = attention_mask.at[original_seq_len:, :].set(False)\n        attention_mask = attention_mask.at[:, original_seq_len:].set(False)\n\n        # Handle causal mask for cached decoder self-attention (from nnx.MultiHeadAttention)\n        if mask_B111 is not None:\n            mask_B111 = _merge_batch_dims(mask_B111)\n            # We need to broadcast T and S dimensions to target_seq_len since cudnn attention strictly checks the mask shape\n            # https://github.com/jax-ml/jax/issues/28974\n            # https://github.com/jax-ml/jax/blob/08c7677393672ccb85c10f1ed0bd506905c3c994/jax/_src/cudnn/fused_attention_stablehlo.py#L1830\n            # https://github.com/jax-ml/jax/blob/08c7677393672ccb85c10f1ed0bd506905c3c994/jax/_src/cudnn/fused_attention_stablehlo.py#L337\n            mask_B1QK = einops.repeat(mask_B111, ""... 1 1 -> ... t s"", t=T, s=T)\n            mask_B1QK = mask_B111.astype(jnp.bool)\n        else:\n            mask_11QK = attention_mask[jnp.newaxis, jnp.newaxis, :, :]\n            mask_B1QK = jnp.broadcast_to(mask_11QK, (B, 1, T, T))\n\n        bias_4d = _pad(_merge_batch_dims(bias)) if bias is not None else None\n\n        # NOTE: jax.nn.dot_product_attention does not support dropout\n        output_4d = jax.nn.dot_product_attention(\n            query=query_BTHD,\n            key=key_BTHD,\n            value=value_BTHD,\n            bias=bias_4d,\n            mask=mask_B1QK,\n            implementation=implementation,\n            is_causal=is_causal,\n        )\n        return output_4d[..., :original_seq_len, :, :].reshape(original_shape)\n\n    return attention_fn\n",python,tab
+6,12453,"utils/nn.py",14603,0,"",python,selection_mouse
+7,12683,"utils/nn.py",0,0,"",python,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-3f97be60-a08c-456f-bfdc-f242b847ea0f1763659243291-2025_11_20-18.21.03.400/source.csv

@@ -0,0 +1,7 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"slurm/dev/franz/berlin/crowd-pilot/nemo/generate_jsonl_nemo_dataset_full.sh",0,0,"#!/bin/bash\n\nset -uex\n\nOUTPUT_DIR=""/fast/project/HFMI_SynergyUnit/tab_model/data/nemo_hf_part_jsonl_full/""\nCSV_ROOT=""/fast/project/HFMI_SynergyUnit/tab_model/data/hf_part_csv/""\n\n# rough estimate of characters per token\nTARGET_CHARS_PER_CONVERSATION=99999999999999999\n\nuv run crowd_pilot/serialize_dataset_nemo_json.py --csv_root=$CSV_ROOT --output_dir=$OUTPUT_DIR --target_chars_per_conversation=$TARGET_CHARS_PER_CONVERSATION --min_session_turns=10",shellscript,tab
+2,33,"slurm/dev/franz/berlin/crowd-pilot/nemo/generate_jsonl_nemo_dataset_full.sh",265,0,"",shellscript,selection_command
+3,167,"slurm/dev/franz/berlin/crowd-pilot/nemo/generate_jsonl_nemo_dataset_full.sh",217,0,"",shellscript,selection_command
+4,216,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"6:21:03 PM [info] Activating crowd-code\n6:21:03 PM [info] Recording started\n6:21:03 PM [info] Initializing git provider using file system watchers...\n6:21:03 PM [info] Git repository found\n6:21:03 PM [info] Git provider initialized successfully\n",Log,tab
+5,1059,"extension-output-pdoom-org.crowd-code-#1-crowd-code",245,0,"6:21:03 PM [info] Initial git state: [object Object]\n",Log,content
+6,1257,"slurm/dev/franz/berlin/crowd-pilot/nemo/generate_jsonl_nemo_dataset_full.sh",0,0,"",shellscript,tab

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-466daf10-f1c4-4543-a0a0-550adf9c52e21755292882142-2025_08_15-23.21.44.913/source.csv

@@ -0,0 +1,10 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"train_lam.py",0,0,"from dataclasses import dataclass, field\nimport os\nfrom typing import cast\n\nimport einops\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom models.lam import LatentActionModel\nfrom utils.dataloader import get_dataloader\nfrom utils.lr_utils import get_lr_schedule\nfrom utils.parameter_utils import count_parameters_by_component\n\njax.config.update(""jax_debug_nans"", True)\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    batch_size: int = 36\n    vq_beta: float = 0.25\n    init_lr: float = 0.0\n    max_lr: float = 3e-5\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    warmup_steps: int = 5000\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    vq_reset_thresh: int = 50\n    # LAM\n    model_dim: int = 512\n    ffn_dim: int = 2048\n    latent_dim: int = 32\n    num_latents: int = 6\n    patch_size: int = 16\n    num_blocks: int = 4\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.0\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_lam""\n    tags: list[str] = field(default_factory=lambda: [""lam""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n    log_checkpoint_keep_period: int = 20000\n    wandb_id: str = """"\n    use_flash_attention: bool = True\n\n\nargs = tyro.cli(Args)\n\n\ndef lam_loss_fn(\n    model: LatentActionModel, inputs: dict\n) -> tuple[jax.Array, tuple[jax.Array, jax.Array, dict]]:\n    # --- Compute loss ---\n    gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n    inputs[""videos""] = gt.astype(args.dtype)\n    model.train()\n    outputs = model(inputs, training=True)\n    outputs[""recon""] = outputs[""recon""].astype(jnp.float32)\n    gt_future_frames = gt[:, 1:]\n    mse = jnp.square(gt_future_frames - outputs[""recon""]).mean()\n    q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n    commitment_loss = jnp.square(\n        outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n    ).mean()\n    loss = mse + q_loss + args.vq_beta * commitment_loss\n\n    # --- Compute validation metrics ---\n    gt = gt_future_frames.clip(0, 1).reshape(-1, *gt_future_frames.shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = jnp.asarray(pix.psnr(gt, recon)).mean()\n    ssim = jnp.asarray(pix.ssim(gt, recon)).mean()\n    count_fn = jax.vmap(lambda i: (outputs[""indices""] == i).sum())\n    index_counts = count_fn(jnp.arange(args.num_latents))\n    metrics = dict(\n        loss=loss,\n        mse=mse,\n        q_loss=q_loss,\n        commitment_loss=commitment_loss,\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage=(index_counts != 0).mean(),\n    )\n    return loss, (outputs[""recon""], index_counts, metrics)\n\n\n@nnx.jit\ndef train_step(\n    lam: LatentActionModel,\n    optimizer: nnx.Optimizer,\n    inputs: dict,\n    action_last_active: jax.Array,\n    rng: jax.Array,\n) -> tuple[jax.Array, jax.Array, jax.Array, dict]:\n    def loss_fn(\n        model: LatentActionModel,\n    ) -> tuple[jax.Array, tuple[jax.Array, jax.Array, dict]]:\n        return lam_loss_fn(model, inputs)\n\n    # --- Update model ---\n    (loss, (recon, idx_counts, metrics)), grads = nnx.value_and_grad(\n        loss_fn, has_aux=True\n    )(lam)\n    optimizer.update(grads)\n\n    # --- Reset inactive latent actions ---\n    codebook = lam.vq.codebook\n    num_codes = len(codebook)\n    active_codes = idx_counts != 0.0\n    action_last_active = jnp.where(active_codes, 0, action_last_active + 1)\n    p_code = active_codes / active_codes.sum()\n    reset_idxs = jax.random.choice(rng, num_codes, shape=(num_codes,), p=p_code)\n    do_reset = action_last_active >= args.vq_reset_thresh\n    new_codebook = jnp.where(\n        jnp.expand_dims(do_reset, -1), codebook[reset_idxs], codebook.value\n    )\n    lam.vq.codebook.value = new_codebook\n    action_last_active = jnp.where(do_reset, 0, action_last_active)\n    return loss, recon, action_last_active, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    lam = LatentActionModel(\n        in_dim=args.image_channels,\n        model_dim=args.model_dim,\n        ffn_dim=args.ffn_dim,\n        latent_dim=args.latent_dim,\n        num_latents=args.num_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.num_blocks,\n        num_heads=args.num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.codebook_dropout,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        rngs=rngs,\n    )\n\n    # Count parameters\n    _, params, _ = nnx.split(lam, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.dtype,\n    )\n    optimizer = nnx.Optimizer(lam, tx)\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n    # --- Initialize checkpoint manager ---\n    step = 0\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n\n    # --- Create DataLoaderIterator from dataloader ---\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n\n    # --- Restore checkpoint ---\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        restored = checkpoint_manager.restore(\n            checkpoint_manager.latest_step(),\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),  # type: ignore\n            ),\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        grain_iterator = restored[""dataloader_state""]\n        step = checkpoint_manager.latest_step() or 0\n        print(f""Restored dataloader and model state from step {step}"")\n\n    # --- TRAIN LOOP ---\n    dataloader = (\n        jax.make_array_from_process_local_data(videos_sharding, elem)\n        for elem in grain_iterator\n    )\n    print(f""Starting training from step {step}..."")\n    action_last_active = jnp.zeros(args.num_latents, dtype=jnp.int32)\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            rng, _rng = jax.random.split(rng)\n\n            inputs = dict(videos=videos, rng=_rng)\n            rng, _rng = jax.random.split(rng)\n            loss, recon, action_last_active, metrics = train_step(\n                lam, optimizer, inputs, action_last_active, _rng\n            )\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0, 1:].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[0])),\n                            recon=wandb.Image(np.asarray(recon_seq[0])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            grain_iterator  # type: ignore\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n",python,tab
+2,1332,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"11:21:44 PM [info] Activating crowd-code\n11:21:44 PM [info] Recording started\n11:21:44 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,1516,"extension-output-pdoom-org.crowd-code-#1-crowd-code",153,0,"11:21:45 PM [info] Git repository found\n11:21:45 PM [info] Git provider initialized successfully\n11:21:45 PM [info] Initial git state: [object Object]\n",Log,content
+4,2562,"TERMINAL",0,0,"",,terminal_focus
+5,2564,"train_lam.py",0,0,"",python,tab
+6,3273,"TERMINAL",0,0,"source /home/franz.srambical/jafar/.venv/bin/activate",,terminal_command
+7,3290,"TERMINAL",0,0,"]633;C]0;franz.srambical@hai-login1:~/jafar",,terminal_output
+8,6082,"TERMINAL",0,0,"squeue",,terminal_command
+9,6090,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  20008 alfred.ngu  standard     1    32  R 2025-08-15T22:08:34 2025-08-15T22:08:36  1:13:14 1-00:00:00 hai006\r\n  20007 alfred.ngu  standard     1    32  R 2025-08-15T22:08:27 2025-08-15T22:08:27  1:13:23 1-00:00:00 hai004\r\n  20005 alfred.ngu  standard     1    32  R 2025-08-15T16:31:14 2025-08-15T16:31:15  6:50:35 1-00:00:00 hai003\r\n  20004 alfred.ngu  standard     1    32  R 2025-08-15T16:21:25 2025-08-15T16:21:25  7:00:25 1-00:00:00 hai002\r\n  19590 nishant.ku  standard     3   192  R 2025-08-15T11:26:40 2025-08-15T11:26:43 11:55:07 1-00:00:00 hai[005,007-008]\r\n]0;franz.srambical@hai-login1:~/jafar",,terminal_output

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-51d6748b-9efc-4fba-a5fd-d9c6e1b9004f1765645139666-2025_12_13-17.59.18.619/source.csv

@@ -0,0 +1,53 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"crates/cli/src/main.rs",0,0,"//! CLI tool for serializing crowd-pilot IDE interaction data.\n//!\n//! This tool processes CSV session files and outputs JSONL format suitable for\n//! NeMo SFT training. It uses an embedded Python interpreter to load HuggingFace\n//! tokenizers for accurate token counting.\n\nuse std::path::PathBuf;\n\nuse clap::Parser;\nuse pyo3::prelude::*;\nuse pyo3::types::PyModule;\n\nuse crowd_pilot_serializer_core::{\n    pipeline::{PipelineConfig, PipelineResult},\n    process_all_sessions, write_jsonl_output, Tokenizer,\n};\n\n/// Serialize crowd-pilot CSV sessions to NeMo JSONL format.\n#[derive(Parser, Debug)]\n#[command(name = ""crowd-pilot-serialize"")]\n#[command(author, version, about, long_about = None)]\nstruct Args {\n    /// Root directory containing CSV session files\n    #[arg(long)]\n    csv_root: PathBuf,\n\n    /// Output directory for JSONL files\n    #[arg(long)]\n    output_dir: PathBuf,\n\n    /// HuggingFace tokenizer model name or path\n    #[arg(long)]\n    tokenizer: String,\n\n    /// Maximum tokens per conversation chunk\n    #[arg(long, default_value = ""8192"")]\n    max_tokens_per_conversation: usize,\n\n    /// Maximum tokens per message\n    #[arg(long, default_value = ""2048"")]\n    max_tokens_per_message: usize,\n\n    /// Minimum messages required to keep a conversation\n    #[arg(long, default_value = ""5"")]\n    min_conversation_messages: usize,\n\n    /// Viewport radius (lines above/below cursor)\n    #[arg(long, default_value = ""10"")]\n    viewport_radius: usize,\n\n    /// Coalesce radius for grouping nearby edits\n    #[arg(long, default_value = ""5"")]\n    coalesce_radius: usize,\n\n    /// Fraction of sessions for validation (0.0-1.0)\n    #[arg(long, default_value = ""0.1"")]\n    val_ratio: f64,\n\n    /// Custom system prompt (optional)\n    #[arg(long)]\n    system_prompt: Option<String>,\n}\n\nconst DEFAULT_SYSTEM_PROMPT: &str = r#""You are a helpful assistant that can interact multiple times with a computer shell to solve programming tasks.\nYour response must contain exactly ONE bash code block with ONE command (or commands connected with && or ||).\n\nFormat your response as shown in <format_example>.\n\n<format_example>\n```bash\nyour_command_here\n```\n</format_example>\n\nFailure to follow these rules will cause your response to be rejected.""#;\n\n/// Wrapper around Python tokenizer for exact token counting and truncation.\nstruct PythonTokenizer {\n    tokenizer: Py<PyAny>,\n}\n\nimpl PythonTokenizer {\n    /// Load a HuggingFace tokenizer.\n    fn load(model_name: &str) -> PyResult<Self> {\n        Python::with_gil(|py| {\n            let transformers = PyModule::import(py, ""transformers"")?;\n            let auto_tokenizer = transformers.getattr(""AutoTokenizer"")?;\n            let tokenizer = auto_tokenizer.call_method1(""from_pretrained"", (model_name,))?;\n            Ok(Self {\n                tokenizer: tokenizer.into(),\n            })\n        })\n    }\n}\n\nimpl Tokenizer for PythonTokenizer {\n    fn count_tokens(&self, text: &str) -> usize {\n        Python::with_gil(|py| {\n            let tokenizer = self.tokenizer.as_ref(py);\n            let tokens = tokenizer\n                .call_method1(""encode"", (text,))\n                .expect(""Failed to encode text with tokenizer"");\n            tokens.len().unwrap()\n        })\n    }\n\n    fn truncate_to_max_tokens(&self, text: &str, max_tokens: usize) -> String {\n        Python::with_gil(|py| {\n            let tokenizer = self.tokenizer.as_ref(py);\n            let kwargs = pyo3::types::PyDict::new(py);\n            kwargs.set_item(""max_length"", max_tokens).unwrap();\n            kwargs.set_item(""truncation"", true).unwrap();\n            \n            let tokens = tokenizer\n                .call_method(""encode"", (text,), Some(kwargs))\n                .expect(""Failed to encode text with tokenizer"");\n            \n            tokenizer\n                .call_method1(""decode"", (tokens,))\n                .expect(""Failed to decode tokens"")\n                .extract()\n                .unwrap()\n        })\n    }\n}\n\nfn main() -> Result<(), Box<dyn std::error::Error>> {\n    let args = Args::parse();\n\n    println!(""Loading tokenizer from {}..."", args.tokenizer);\n    let tokenizer = PythonTokenizer::load(&args.tokenizer)?;\n\n    let config = PipelineConfig {\n        max_tokens_per_conversation: args.max_tokens_per_conversation,\n        max_tokens_per_message: args.max_tokens_per_message,\n        min_conversation_messages: args.min_conversation_messages,\n        viewport_radius: args.viewport_radius,\n        coalesce_radius: args.coalesce_radius,\n        val_ratio: args.val_ratio,\n    };\n\n    println!(""Processing CSV files from {:?}..."", args.csv_root);\n    let session_results = process_all_sessions(\n        &args.csv_root,\n        &tokenizer,\n        &config,\n    )?;\n\n    let total_sessions = session_results.len();\n    println!(""Processed {} sessions"", total_sessions);\n\n    let system_prompt = args.system_prompt.as_deref().unwrap_or(DEFAULT_SYSTEM_PROMPT);\n\n    println!(""Writing output to {:?}..."", args.output_dir);\n    let result: PipelineResult = write_jsonl_output(\n        session_results,\n        &args.output_dir,\n        args.val_ratio,\n        system_prompt,\n    )?;\n\n    let metadata_path = args.output_dir.join(""metadata.json"");\n    let metadata = serde_json::json!({\n        ""config"": {\n            ""csv_root"": args.csv_root.to_string_lossy(),\n            ""output_dir"": args.output_dir.to_string_lossy(),\n            ""tokenizer"": args.tokenizer,\n            ""max_tokens_per_conversation"": args.max_tokens_per_conversation,\n            ""max_tokens_per_message"": args.max_tokens_per_message,\n            ""min_conversation_messages"": args.min_conversation_messages,\n            ""viewport_radius"": args.viewport_radius,\n            ""coalesce_radius"": args.coalesce_radius,\n            ""val_ratio"": args.val_ratio,\n        },\n        ""counts"": {\n            ""total_sessions"": result.total_sessions,\n            ""total_conversations"": result.total_conversations,\n            ""train_conversations"": result.train_conversations,\n            ""val_conversations"": result.val_conversations,\n        },\n        ""stats"": {\n            ""total_messages"": result.total_messages,\n            ""total_tokens"": result.total_tokens,\n            ""avg_messages_per_conversation"": if result.total_conversations > 0 {\n                result.total_messages as f64 / result.total_conversations as f64\n            } else {\n                0.0\n            },\n            ""avg_tokens_per_conversation"": if result.total_conversations > 0 {\n                result.total_tokens as f64 / result.total_conversations as f64\n            } else {\n                0.0\n            },\n        },\n        ""files"": {\n            ""train_path"": args.output_dir.join(""training.jsonl"").to_string_lossy(),\n            ""val_path"": args.output_dir.join(""validation.jsonl"").to_string_lossy(),\n        },\n    });\n    std::fs::write(&metadata_path, serde_json::to_string_pretty(&metadata)?)?;\n\n    println!(""\n[summary]"");\n    println!(""  Total sessions processed: {}"", result.total_sessions);\n    println!(""  Train conversations: {}"", result.train_conversations);\n    println!(""  Val conversations: {}"", result.val_conversations);\n    println!(""  Total messages: {}"", result.total_messages);\n    println!(""  Total tokens: {}"", result.total_tokens);\n    println!(""  Output: {:?}/{{training,validation}}.jsonl"", args.output_dir);\n    println!(""  Metadata: {:?}"", metadata_path);\n\n    Ok(())\n}\n\n",rust,tab
+2,335,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"5:59:18 PM [info] Activating crowd-code\n5:59:18 PM [info] Recording started\n5:59:18 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,662,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"5:59:18 PM [info] Git repository found\n5:59:18 PM [info] Git provider initialized successfully\n5:59:18 PM [info] Initial git state: [object Object]\n",Log,content
+4,3455,"TERMINAL",0,0,"",,terminal_focus
+5,3456,"crates/cli/src/main.rs",0,0,"",rust,tab
+6,19365,"TERMINAL",0,0,"cargo run --bin crowd-pilot-serialize -- --help",,terminal_command
+7,19415,"TERMINAL",0,0,"]633;C",,terminal_output
+8,19980,"TERMINAL",0,0,"   Compiling pyo3-build-config v0.20.3\r\n    Building [======================>    ] 73/84: pyo3-build-config(build)                                                               \r",,terminal_output
+9,20326,"TERMINAL",0,0,"    Building [======================>    ] 74/84: pyo3-build-config                                                                      \r",,terminal_output
+10,20374,"TERMINAL",0,0,"   Compiling pyo3-macros-backend v0.20.3\r\n    Building [======================>    ] 74/84: pyo3-build-config, pyo3-macros-backend                                                 \r",,terminal_output
+11,20558,"TERMINAL",0,0,"   Compiling pyo3-ffi v0.20.3\r\n    Building [=======================>   ] 75/84: pyo3-ffi(build.rs), pyo3-macros-backend                                                \r   Compiling pyo3 v0.20.3\r\n    Building [=======================>   ] 75/84: pyo3-ffi(build.rs), pyo3-macros-backend, pyo3(build.rs)                                \r",,terminal_output
+12,20901,"TERMINAL",0,0,"    Building [=======================>   ] 76/84: pyo3-ffi(build.rs), pyo3-macros-backend                                                \r    Building [=======================>   ] 77/84: pyo3-macros-backend, pyo3-ffi(build)                                                   \r",,terminal_output
+13,21039,"TERMINAL",0,0,"    Building [========================>  ] 78/84: pyo3-ffi, pyo3-macros-backend, pyo3(build)                                             \r    Building [========================>  ] 79/84: pyo3-ffi, pyo3-macros-backend                                                          \r",,terminal_output
+14,21355,"TERMINAL",0,0,"    Building [========================>  ] 80/84: pyo3-macros-backend                                                                    \r",,terminal_output
+15,21638,"TERMINAL",0,0,"   Compiling pyo3-macros v0.20.3\r\n    Building [=========================> ] 81/84: pyo3-macros                                                                            \r",,terminal_output
+16,21849,"TERMINAL",0,0,"    Building [=========================> ] 82/84: pyo3                                                                                   \r",,terminal_output
+17,23797,"TERMINAL",0,0,"   Compiling crowd-pilot-serialize v0.1.0 (/fast/home/franz.srambical/crowd-pilot-serializer/crates/cli)\r\n    Building [=========================> ] 83/84: crowd-pilot-serialize(bin)                                                             \r",,terminal_output
+18,25819,"TERMINAL",0,0,"    Finished ]8;;https://doc.rust-lang.org/cargo/reference/profiles.html#default-profiles\`dev` profile [unoptimized + debuginfo]]8;;\ target(s) in 6.31s\r\n     Running `target/debug/crowd-pilot-serialize --help`\r\ntarget/debug/crowd-pilot-serialize: error while loading shared libraries: libpython3.12.so.1.0: cannot open shared object file: No such file or directory\r\n]0;franz.srambical@hai-login2:~/crowd-pilot-serializer",,terminal_output
+19,79950,"TERMINAL",0,0,"python3 --version",,terminal_command
+20,80680,"TERMINAL",0,0,"python3 --versionn",,terminal_command
+21,82077,"TERMINAL",0,0,"python3 --version",,terminal_command
+22,106622,"TERMINAL",0,0,"python3 -c ""import sysconfig; print(sysconfig.get_config_var('LIBDIR'))\n""",,terminal_command
+23,106623,"TERMINAL",0,0,"]633;C/opt/miniforge3/lib\r\n]0;franz.srambical@hai-login2:~/crowd-pilot-serializer",,terminal_output
+24,213021,"TERMINAL",0,0,"uv venv",,terminal_command
+25,213071,"TERMINAL",0,0,"]633;C",,terminal_output
+26,213264,"TERMINAL",0,0,"Using CPython 3.13.5\r\nCreating virtual environment at: .venv\r\nActivate with: source .venv/bin/activate\r\n]0;franz.srambical@hai-login2:~/crowd-pilot-serializer",,terminal_output
+27,222429,"TERMINAL",0,0,"source .venv/bin/activate",,terminal_command
+28,267603,"TERMINAL",0,0,"cargo run --bin crowd-pilot-serialize -- --help",,terminal_command
+29,267658,"TERMINAL",0,0,"]633;C",,terminal_output
+30,268298,"TERMINAL",0,0,"   Compiling pyo3-build-config v0.20.3\r\n    Building [======================>    ] 73/84: pyo3-build-config(build)                                                               \r",,terminal_output
+31,269082,"TERMINAL",0,0,"    Building [======================>    ] 74/84: pyo3-build-config                                                                      \r",,terminal_output
+32,269324,"TERMINAL",0,0,"   Compiling pyo3-macros-backend v0.20.3\r\n    Building [======================>    ] 74/84: pyo3-macros-backend, pyo3-build-config                                                 \r",,terminal_output
+33,269500,"TERMINAL",0,0,"   Compiling pyo3-ffi v0.20.3\r\n    Building [=======================>   ] 75/84: pyo3-ffi(build.rs), pyo3-macros-backend                                                \r   Compiling pyo3 v0.20.3\r\n    Building [=======================>   ] 75/84: pyo3-ffi(build.rs), pyo3-macros-backend, pyo3(build.rs)                                \r",,terminal_output
+34,269801,"TERMINAL",0,0,"    Building [=======================>   ] 76/84: pyo3-ffi(build.rs), pyo3-macros-backend                                                \r    Building [=======================>   ] 77/84: pyo3-macros-backend, pyo3-ffi(build)                                                   \rerror: failed to run custom build command for `pyo3-ffi v0.20.3`\r\n\r\nCaused by:\r\n  process didn't exit successfully: `/fast/home/franz.srambical/crowd-pilot-serializer/target/debug/build/pyo3-ffi-c29aa9f4ee129c43/build-script-build` (exit status: 1)\r\n  --- stdout\r\n  cargo:rerun-if-env-changed=PYO3_CROSS\r\n  cargo:rerun-if-env-changed=PYO3_CROSS_LIB_DIR\r\n  cargo:rerun-if-env-changed=PYO3_CROSS_PYTHON_VERSION\r\n  cargo:rerun-if-env-changed=PYO3_CROSS_PYTHON_IMPLEMENTATION\r\n  cargo:rerun-if-env-changed=PYO3_PRINT_CONFIG\r\n  cargo:rerun-if-env-changed=PYO3_USE_ABI3_FORWARD_COMPATIBILITY\r\n\r\n  --- stderr\r\n  error: the configured Python interpreter version (3.13) is newer than PyO3's maximum supported version (3.12)\r\n  = help: please check if an updated version of PyO3 is available. Current version: 0.20.3\r\n  = help: set PYO3_USE_ABI3_FORWARD_COMPATIBILITY=1 to suppress this check and build anyway using the stable ABI\r\nwarning: build failed, waiting for other jobs to finish...\r\n    Building [========================>  ] 78/84: pyo3-macros-backend                                                                    \r",,terminal_output
+35,270412,"TERMINAL",0,0,"]0;franz.srambical@hai-login2:~/crowd-pilot-serializer",,terminal_output
+36,312052,"TERMINAL",0,0,"deactivate",,terminal_command
+37,312232,"TERMINAL",0,0,"",,terminal_command
+38,326164,"TERMINAL",0,0,"rm -rf .venv/",,terminal_command
+39,326167,"TERMINAL",0,0,"]633;C]0;franz.srambical@hai-login2:~/crowd-pilot-serializer",,terminal_output
+40,341417,"TERMINAL",0,0,"uv venv --python 3.12",,terminal_command
+41,341419,"TERMINAL",0,0,"]633;CUsing CPython 3.12.9 interpreter at: /opt/miniforge3/bin/python3.12\r\nCreating virtual environment at: .venv\r\nActivate with: source .venv/bin/activate\r\n]0;franz.srambical@hai-login2:~/crowd-pilot-serializer",,terminal_output
+42,344601,"TERMINAL",0,0,"source .venv/bin/activate",,terminal_command
+43,348037,"TERMINAL",0,0,"cargo run --bin crowd-pilot-serialize -- --help",,terminal_command
+44,348087,"TERMINAL",0,0,"]633;C",,terminal_output
+45,348628,"TERMINAL",0,0,"   Compiling pyo3-ffi v0.20.3\r\n   Compiling pyo3-macros v0.20.3\r\n    Building [========================>  ] 78/84: pyo3-ffi(build), pyo3-macros                                                           \r",,terminal_output
+46,348820,"TERMINAL",0,0,"error: failed to run custom build command for `pyo3-ffi v0.20.3`\r\n\r\nCaused by:\r\n  process didn't exit successfully: `/fast/home/franz.srambical/crowd-pilot-serializer/target/debug/build/pyo3-ffi-c29aa9f4ee129c43/build-script-build` (exit status: 1)\r\n  --- stdout\r\n  cargo:rerun-if-env-changed=PYO3_CROSS\r\n  cargo:rerun-if-env-changed=PYO3_CROSS_LIB_DIR\r\n  cargo:rerun-if-env-changed=PYO3_CROSS_PYTHON_VERSION\r\n  cargo:rerun-if-env-changed=PYO3_CROSS_PYTHON_IMPLEMENTATION\r\n  cargo:rerun-if-env-changed=PYO3_PRINT_CONFIG\r\n  cargo:rerun-if-env-changed=PYO3_USE_ABI3_FORWARD_COMPATIBILITY\r\n\r\n  --- stderr\r\n  error: the configured Python interpreter version (3.13) is newer than PyO3's maximum supported version (3.12)\r\n  = help: please check if an updated version of PyO3 is available. Current version: 0.20.3\r\n  = help: set PYO3_USE_ABI3_FORWARD_COMPATIBILITY=1 to suppress this check and build anyway using the stable ABI\r\nwarning: build failed, waiting for other jobs to finish...\r\n    Building [========================>  ] 79/84: pyo3-macros                                                                            \r",,terminal_output
+47,349008,"TERMINAL",0,0,"]0;franz.srambical@hai-login2:~/crowd-pilot-serializer",,terminal_output
+48,399487,"TERMINAL",0,0,"python --version",,terminal_command
+49,399488,"TERMINAL",0,0,"]633;CPython 3.12.9\r\n]0;franz.srambical@hai-login2:~/crowd-pilot-serializer",,terminal_output
+50,400945,"TERMINAL",0,0,"which python",,terminal_command
+51,401013,"TERMINAL",0,0,"]633;C",,terminal_output
+52,401018,"TERMINAL",0,0,"/fast/home/franz.srambical/crowd-pilot-serializer/.venv/bin/python\r\n]0;franz.srambical@hai-login2:~/crowd-pilot-serializer",,terminal_output

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-55481ea8-1b17-4c2d-a934-2a959c08e0db1754989165246-2025_08_12-10.59.35.36/source.csv

@@ -0,0 +1,269 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"train_dynamics.py",0,0,"from dataclasses import dataclass, field\nimport os\nfrom typing import cast\n\nimport einops\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom genie import Genie, restore_genie_components\nfrom utils.dataloader import get_dataloader\nfrom utils.lr_utils import get_lr_schedule\nfrom utils.parameter_utils import count_parameters_by_component\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    batch_size: int = 36\n    init_lr: float = 0.0\n    max_lr: float = 3e-5\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    warmup_steps: int = 5000\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    # Tokenizer\n    tokenizer_dim: int = 512\n    tokenizer_ffn_dim: int = 2048\n    latent_patch_dim: int = 32\n    num_patch_latents: int = 1024\n    patch_size: int = 4\n    tokenizer_num_blocks: int = 4\n    tokenizer_num_heads: int = 8\n    tokenizer_checkpoint: str = """"\n    # LAM\n    lam_dim: int = 512\n    lam_ffn_dim: int = 2048\n    latent_action_dim: int = 32\n    num_latent_actions: int = 6\n    lam_patch_size: int = 16\n    lam_num_blocks: int = 4\n    lam_num_heads: int = 8\n    lam_checkpoint: str = """"\n    # Dynamics\n    dyna_type: str = ""maskgit""  # supported options: maskgit, causal\n    dyna_dim: int = 512\n    dyna_ffn_dim: int = 2048\n    dyna_num_blocks: int = 6\n    dyna_num_heads: int = 8\n    dropout: float = 0.0\n    mask_limit: float = 0.5\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    use_flash_attention: bool = True\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_dynamics""\n    tags: list[str] = field(default_factory=lambda: [""dynamics""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 25000\n    log_checkpoint_keep_period: int = 20000\n    log_gradients: bool = False\n    wandb_id: str = """"\n\n\nargs = tyro.cli(Args)\n\n\ndef dynamics_loss_fn(\n    model: Genie, inputs: dict\n) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n    """"""Compute masked dynamics loss""""""\n    # gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n    # inputs[""videos""] = gt.astype(args.dtype)\n    model.train()\n    outputs = model(inputs, training=True)\n    mask = outputs[""mask""]\n    outputs[""token_logits""] = outputs[""token_logits""].astype(jnp.float32)\n    ce_loss = optax.softmax_cross_entropy_with_integer_labels(\n        outputs[""token_logits""], outputs[""video_tokens""]\n    )\n    ce_loss = (mask * ce_loss).sum() / mask.sum()\n    acc = outputs[""token_logits""].argmax(-1) == outputs[""video_tokens""]\n    acc = (mask * acc).sum() / mask.sum()\n    select_probs = jax.nn.softmax(outputs[""token_logits""])\n    # gt = gt.clip(0, 1).reshape(-1, *gt.shape[2:])\n    # recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    # psnr = jnp.asarray(pix.psnr(gt, recon)).mean()\n    # ssim = jnp.asarray(pix.ssim(gt, recon)).mean()\n    # _, index_counts_lam = jnp.unique_counts(\n    #     jnp.ravel(outputs[""lam_indices""]), size=args.num_latent_actions, fill_value=0\n    # )\n    _, index_counts_tokenizer = jnp.unique_counts(\n        jnp.ravel(outputs[""video_tokens""]), size=args.num_patch_latents, fill_value=0\n    )\n    # codebook_usage_lam = (index_counts_lam != 0).mean()\n    codebook_usage_tokenizer = (index_counts_tokenizer != 0).mean()\n    metrics = dict(\n        cross_entropy_loss=ce_loss,\n        masked_token_accuracy=acc,\n        select_logit=outputs[""token_logits""].max(-1).mean(),\n        select_p=select_probs.max(-1).mean(),\n        entropy=jax.scipy.special.entr(select_probs).sum(-1).mean(),\n        # psnr=psnr,\n        # ssim=ssim,\n        # codebook_usage_lam=codebook_usage_lam,\n        codebook_usage_tokenizer=codebook_usage_tokenizer,\n    )\n    return ce_loss, (None, metrics)\n\n\n@nnx.jit\ndef train_step(\n    model: Genie, optimizer: nnx.Optimizer, inputs: dict\n) -> tuple[jax.Array, jax.Array, dict]:\n    """"""Update state and compute metrics""""""\n\n    def loss_fn(model: Genie) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n        return dynamics_loss_fn(model, inputs)\n\n    (loss, (recon, metrics)), grads = nnx.value_and_grad(loss_fn, has_aux=True)(model)\n    optimizer.update(grads)\n    if args.log_gradients:\n        metrics[""gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""dynamics""]\n        )\n    return loss, recon, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    genie = Genie(\n        # Tokenizer\n        in_dim=args.image_channels,\n        tokenizer_dim=args.tokenizer_dim,\n        tokenizer_ffn_dim=args.tokenizer_ffn_dim,\n        latent_patch_dim=args.latent_patch_dim,\n        num_patch_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        tokenizer_num_blocks=args.tokenizer_num_blocks,\n        tokenizer_num_heads=args.tokenizer_num_heads,\n        # LAM\n        lam_dim=args.lam_dim,\n        lam_ffn_dim=args.lam_ffn_dim,\n        latent_action_dim=args.latent_action_dim,\n        num_latent_actions=args.num_latent_actions,\n        lam_patch_size=args.lam_patch_size,\n        lam_num_blocks=args.lam_num_blocks,\n        lam_num_heads=args.lam_num_heads,\n        lam_co_train=not args.lam_checkpoint,\n        # Dynamics\n        dyna_type=args.dyna_type,\n        dyna_dim=args.dyna_dim,\n        dyna_ffn_dim=args.dyna_ffn_dim,\n        dyna_num_blocks=args.dyna_num_blocks,\n        dyna_num_heads=args.dyna_num_heads,\n        dropout=args.dropout,\n        mask_limit=args.mask_limit,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        decode=False,\n        rngs=rngs,\n    )\n\n    _, params, _ = nnx.split(genie, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.dtype,\n    )\n    optimizer = nnx.Optimizer(genie, tx)\n    del genie\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n    # --- Initialize checkpoint manager ---\n    step = 0\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n\n    # --- Create DataLoaderIterator from dataloader ---\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n\n    # --- Restore checkpoint ---\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        restored = checkpoint_manager.restore(\n            checkpoint_manager.latest_step(),\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),  # type: ignore\n            ),\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        grain_iterator = restored[""dataloader_state""]\n        step = checkpoint_manager.latest_step() or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    else:\n        # Restore from pre-trained tokenizer (and LAM)\n        optimizer = restore_genie_components(optimizer, replicated_sharding, rng, args)\n        # NOTE: We have to remove the (unused) tokenizer vq dropout due flax.nnx lazily initializing modules.\n        # Specifically, the first dynamics model checkpoint will contain the vq dropout module,\n        # but the first full restore will fail due to nnx not initializing the module when\n        # dropout is set to 0.0.\n        del optimizer.model.tokenizer.vq.drop\n\n    # --- TRAIN LOOP ---\n    # dataloader = (\n    #     jax.make_array_from_process_local_data(videos_sharding, elem)\n    #     for elem in grain_iterator\n    # )\n    print(f""Starting training from step {step}..."")\n    while step < args.num_steps:\n        for _ in range(5):\n            # --- Train step ---\n            rng, _rng_mask = jax.random.split(rng, 2)\n            inputs = dict(mask_rng=_rng_mask)\n            loss, recon, metrics = train_step(optimizer.model, optimizer, inputs)\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    pass\n                    # gt_seq = inputs[""videos""][0].astype(jnp.float32) / 255.0\n                    # recon_seq = recon[0].clip(0, 1)\n                    # comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    # comparison_seq = einops.rearrange(\n                    #     comparison_seq * 255, ""t h w c -> h (t w) c""\n                    # )\n                    # if jax.process_index() == 0:\n                    #     log_images = dict(\n                    #         image=wandb.Image(np.asarray(gt_seq[args.seq_len - 1])),\n                    #         recon=wandb.Image(np.asarray(recon_seq[args.seq_len - 1])),\n                    #         true_vs_recon=wandb.Image(\n                    #             np.asarray(comparison_seq.astype(np.uint8))\n                    #         ),\n                    #     )\n                    #     wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            grain_iterator  # type: ignore\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n",python,tab
+2,333,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"10:59:35 AM [info] Activating crowd-code\n10:59:35 AM [info] Recording started\n10:59:35 AM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,506,"extension-output-pdoom-org.crowd-code-#1-crowd-code",153,0,"10:59:35 AM [info] Git repository found\n10:59:35 AM [info] Git provider initialized successfully\n10:59:35 AM [info] Initial git state: [object Object]\n",Log,content
+4,1202,"train_dynamics.py",0,0,"",python,tab
+5,77229,"train_dynamics.py",12272,0,"",python,selection_command
+6,274249,"train_dynamics.py",12271,0,"\n    ",python,content
+7,274462,"train_dynamics.py",12276,0,"w",python,content
+8,274463,"train_dynamics.py",12277,0,"",python,selection_keyboard
+9,274520,"train_dynamics.py",12277,0,"i",python,content
+10,274521,"train_dynamics.py",12278,0,"",python,selection_keyboard
+11,274589,"train_dynamics.py",12278,0,"t",python,content
+12,274590,"train_dynamics.py",12279,0,"",python,selection_keyboard
+13,274763,"train_dynamics.py",12279,0,"h",python,content
+14,274764,"train_dynamics.py",12280,0,"",python,selection_keyboard
+15,275058,"train_dynamics.py",12280,0," ",python,content
+16,275059,"train_dynamics.py",12281,0,"",python,selection_keyboard
+17,275070,"train_dynamics.py",12281,0,"j",python,content
+18,275070,"train_dynamics.py",12282,0,"",python,selection_keyboard
+19,275073,"train_dynamics.py",12282,0,"a",python,content
+20,275073,"train_dynamics.py",12283,0,"",python,selection_keyboard
+21,275134,"train_dynamics.py",12283,0,"x",python,content
+22,275134,"train_dynamics.py",12284,0,"",python,selection_keyboard
+23,275285,"train_dynamics.py",12284,0,".",python,content
+24,275286,"train_dynamics.py",12285,0,"",python,selection_keyboard
+25,275448,"train_dynamics.py",12285,0,"p",python,content
+26,275449,"train_dynamics.py",12286,0,"",python,selection_keyboard
+27,275516,"train_dynamics.py",12286,0,"r",python,content
+28,275516,"train_dynamics.py",12287,0,"",python,selection_keyboard
+29,275710,"train_dynamics.py",12287,0,"o",python,content
+30,275710,"train_dynamics.py",12288,0,"",python,selection_keyboard
+31,275825,"train_dynamics.py",12288,0,"f",python,content
+32,275826,"train_dynamics.py",12289,0,"",python,selection_keyboard
+33,275925,"train_dynamics.py",12289,0,"i",python,content
+34,275925,"train_dynamics.py",12290,0,"",python,selection_keyboard
+35,275985,"train_dynamics.py",12290,0,"l",python,content
+36,275985,"train_dynamics.py",12291,0,"",python,selection_keyboard
+37,276086,"train_dynamics.py",12291,0,"e",python,content
+38,276087,"train_dynamics.py",12292,0,"",python,selection_keyboard
+39,278088,"train_dynamics.py",12292,0,"r",python,content
+40,278089,"train_dynamics.py",12293,0,"",python,selection_keyboard
+41,278222,"train_dynamics.py",12293,0,".",python,content
+42,278222,"train_dynamics.py",12294,0,"",python,selection_keyboard
+43,278356,"train_dynamics.py",12294,0,"t",python,content
+44,278356,"train_dynamics.py",12295,0,"",python,selection_keyboard
+45,278681,"train_dynamics.py",12295,0,"r",python,content
+46,278682,"train_dynamics.py",12296,0,"",python,selection_keyboard
+47,278683,"train_dynamics.py",12296,0,"a",python,content
+48,278683,"train_dynamics.py",12297,0,"",python,selection_keyboard
+49,278716,"train_dynamics.py",12297,0,"c",python,content
+50,278716,"train_dynamics.py",12298,0,"",python,selection_keyboard
+51,278798,"train_dynamics.py",12298,0,"e",python,content
+52,278798,"train_dynamics.py",12299,0,"",python,selection_keyboard
+53,279268,"train_dynamics.py",12299,0,"()",python,content
+54,279269,"train_dynamics.py",12300,0,"",python,selection_keyboard
+55,279505,"train_dynamics.py",12300,0,"""""",python,content
+56,279506,"train_dynamics.py",12301,0,"",python,selection_keyboard
+57,284028,"train_dynamics.py",12301,0,"e",python,content
+58,284028,"train_dynamics.py",12302,0,"",python,selection_keyboard
+59,284306,"train_dynamics.py",12301,1,"",python,content
+60,284357,"train_dynamics.py",12301,0,"t",python,content
+61,284357,"train_dynamics.py",12302,0,"",python,selection_keyboard
+62,284471,"train_dynamics.py",12302,0,"e",python,content
+63,284471,"train_dynamics.py",12303,0,"",python,selection_keyboard
+64,284547,"train_dynamics.py",12303,0,"n",python,content
+65,284547,"train_dynamics.py",12304,0,"",python,selection_keyboard
+66,284708,"train_dynamics.py",12304,0,"s",python,content
+67,284709,"train_dynamics.py",12305,0,"",python,selection_keyboard
+68,284860,"train_dynamics.py",12305,0,"o",python,content
+69,284861,"train_dynamics.py",12306,0,"",python,selection_keyboard
+70,284957,"train_dynamics.py",12306,0,"r",python,content
+71,284957,"train_dynamics.py",12307,0,"",python,selection_keyboard
+72,285119,"train_dynamics.py",12307,0,"b",python,content
+73,285120,"train_dynamics.py",12308,0,"",python,selection_keyboard
+74,285231,"train_dynamics.py",12308,0,"o",python,content
+75,285231,"train_dynamics.py",12309,0,"",python,selection_keyboard
+76,285232,"train_dynamics.py",12309,0,"a",python,content
+77,285233,"train_dynamics.py",12310,0,"",python,selection_keyboard
+78,285526,"train_dynamics.py",12310,0,"r",python,content
+79,285527,"train_dynamics.py",12311,0,"",python,selection_keyboard
+80,285607,"train_dynamics.py",12311,0,"d",python,content
+81,285607,"train_dynamics.py",12312,0,"",python,selection_keyboard
+82,285766,"train_dynamics.py",12311,0,"",python,selection_command
+83,290466,"train_dynamics.py",12314,0,"",python,selection_command
+84,290471,"train_dynamics.py",12314,0,":",python,content
+85,290471,"train_dynamics.py",12315,0,"",python,selection_keyboard
+86,290558,"train_dynamics.py",12315,0,"\n        ",python,content
+87,291116,"train_dynamics.py",12316,8,"",python,content
+88,292911,"train_dynamics.py",12316,1,"",python,content
+89,292918,"train_dynamics.py",12320,0,"",python,selection_command
+90,293224,"train_dynamics.py",12316,32,"    while step < args.num_steps:",python,selection_command
+91,293533,"train_dynamics.py",12316,391,"    while step < args.num_steps:\n        for _ in range(5):\n            # --- Train step ---\n            rng, _rng_mask = jax.random.split(rng, 2)\n            inputs = dict(mask_rng=_rng_mask)\n            loss, recon, metrics = train_step(optimizer.model, optimizer, inputs)\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n",python,selection_command
+92,293689,"train_dynamics.py",12316,2426,"    while step < args.num_steps:\n        for _ in range(5):\n            # --- Train step ---\n            rng, _rng_mask = jax.random.split(rng, 2)\n            inputs = dict(mask_rng=_rng_mask)\n            loss, recon, metrics = train_step(optimizer.model, optimizer, inputs)\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    pass\n                    # gt_seq = inputs[""videos""][0].astype(jnp.float32) / 255.0\n                    # recon_seq = recon[0].clip(0, 1)\n                    # comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    # comparison_seq = einops.rearrange(\n                    #     comparison_seq * 255, ""t h w c -> h (t w) c""\n                    # )\n                    # if jax.process_index() == 0:\n                    #     log_images = dict(\n                    #         image=wandb.Image(np.asarray(gt_seq[args.seq_len - 1])),\n                    #         recon=wandb.Image(np.asarray(recon_seq[args.seq_len - 1])),\n                    #         true_vs_recon=wandb.Image(\n                    #             np.asarray(comparison_seq.astype(np.uint8))\n                    #         ),\n                    #     )\n                    #     wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            grain_iterator  # type: ignore\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n",python,selection_command
+93,293810,"train_dynamics.py",12316,2458,"    while step < args.num_steps:\n        for _ in range(5):\n            # --- Train step ---\n            rng, _rng_mask = jax.random.split(rng, 2)\n            inputs = dict(mask_rng=_rng_mask)\n            loss, recon, metrics = train_step(optimizer.model, optimizer, inputs)\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    pass\n                    # gt_seq = inputs[""videos""][0].astype(jnp.float32) / 255.0\n                    # recon_seq = recon[0].clip(0, 1)\n                    # comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    # comparison_seq = einops.rearrange(\n                    #     comparison_seq * 255, ""t h w c -> h (t w) c""\n                    # )\n                    # if jax.process_index() == 0:\n                    #     log_images = dict(\n                    #         image=wandb.Image(np.asarray(gt_seq[args.seq_len - 1])),\n                    #         recon=wandb.Image(np.asarray(recon_seq[args.seq_len - 1])),\n                    #         true_vs_recon=wandb.Image(\n                    #             np.asarray(comparison_seq.astype(np.uint8))\n                    #         ),\n                    #     )\n                    #     wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            grain_iterator  # type: ignore\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n",python,selection_command
+94,294874,"train_dynamics.py",14743,4,"        ",python,content
+95,294874,"train_dynamics.py",14720,16,"                    ",python,content
+96,294874,"train_dynamics.py",14681,12,"                ",python,content
+97,294874,"train_dynamics.py",14623,16,"                    ",python,content
+98,294874,"train_dynamics.py",14605,16,"                    ",python,content
+99,294874,"train_dynamics.py",14582,20,"                        ",python,content
+100,294874,"train_dynamics.py",14555,24,"                            ",python,content
+101,294874,"train_dynamics.py",14496,28,"                                ",python,content
+102,294874,"train_dynamics.py",14406,24,"                            ",python,content
+103,294874,"train_dynamics.py",14316,24,"                            ",python,content
+104,294874,"train_dynamics.py",14271,20,"                        ",python,content
+105,294874,"train_dynamics.py",14245,20,"                        ",python,content
+106,294874,"train_dynamics.py",14204,16,"                    ",python,content
+107,294874,"train_dynamics.py",14149,16,"                    ",python,content
+108,294874,"train_dynamics.py",14073,12,"                ",python,content
+109,294874,"train_dynamics.py",14037,12,"                ",python,content
+110,294874,"train_dynamics.py",13989,20,"                        ",python,content
+111,294874,"train_dynamics.py",13961,20,"                        ",python,content
+112,294874,"train_dynamics.py",13928,20,"                        ",python,content
+113,294874,"train_dynamics.py",13850,20,"                        ",python,content
+114,294874,"train_dynamics.py",13793,20,"                        ",python,content
+115,294874,"train_dynamics.py",13703,20,"                        ",python,content
+116,294874,"train_dynamics.py",13616,20,"                        ",python,content
+117,294874,"train_dynamics.py",13571,20,"                        ",python,content
+118,294874,"train_dynamics.py",13520,20,"                        ",python,content
+119,294874,"train_dynamics.py",13496,20,"                        ",python,content
+120,294874,"train_dynamics.py",13425,20,"                        ",python,content
+121,294874,"train_dynamics.py",13368,20,"                        ",python,content
+122,294874,"train_dynamics.py",13284,20,"                        ",python,content
+123,294874,"train_dynamics.py",13230,20,"                        ",python,content
+124,294874,"train_dynamics.py",13151,20,"                        ",python,content
+125,294874,"train_dynamics.py",13126,20,"                        ",python,content
+126,294874,"train_dynamics.py",13070,16,"                    ",python,content
+127,294874,"train_dynamics.py",13048,20,"                        ",python,content
+128,294874,"train_dynamics.py",13022,24,"                            ",python,content
+129,294874,"train_dynamics.py",12983,28,"                                ",python,content
+130,294874,"train_dynamics.py",12941,28,"                                ",python,content
+131,294874,"train_dynamics.py",12899,28,"                                ",python,content
+132,294874,"train_dynamics.py",12873,24,"                            ",python,content
+133,294874,"train_dynamics.py",12842,20,"                        ",python,content
+134,294874,"train_dynamics.py",12763,16,"                    ",python,content
+135,294874,"train_dynamics.py",12738,12,"                ",python,content
+136,294874,"train_dynamics.py",12708,12,"                ",python,content
+137,294874,"train_dynamics.py",12685,12,"                ",python,content
+138,294874,"train_dynamics.py",12637,12,"                ",python,content
+139,294874,"train_dynamics.py",12591,12,"                ",python,content
+140,294874,"train_dynamics.py",12509,12,"                ",python,content
+141,294874,"train_dynamics.py",12463,12,"                ",python,content
+142,294874,"train_dynamics.py",12409,12,"                ",python,content
+143,294874,"train_dynamics.py",12376,12,"                ",python,content
+144,294874,"train_dynamics.py",12349,8,"            ",python,content
+145,294874,"train_dynamics.py",12316,4,"        ",python,content
+146,294909,"train_dynamics.py",12324,0,"",python,selection_command
+147,295698,"train_dynamics.py",12280,0,"",python,selection_command
+148,304412,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+149,305759,"TERMINAL",0,0,"",,terminal_focus
+150,305760,"train_dynamics.py",0,0,"",python,tab
+151,307658,"TERMINAL",0,0,"source /home/franz.srambical/jafar/.venv/bin/activate",,terminal_command
+152,318622,"TERMINAL",0,0,"bash experiments/dynamics_grain_tok_restore.sh",,terminal_command
+153,318782,"TERMINAL",0,0,"]633;C",,terminal_output
+154,325835,"TERMINAL",0,0,"2025-08-12 11:05:00.706052: E external/xla/xla/stream_executor/cuda/cuda_platform.cc:51] failed call to cuInit: INTERNAL: CUDA error: Failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected\r\nTraceback (most recent call last):\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/jax/_src/xla_bridge.py"", line 742, in backends\r\n    backend = _init_backend(platform)\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/jax/_src/xla_bridge.py"", line 828, in _init_backend\r\n    backend = registration.factory()\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/jax/_src/xla_bridge.py"", line 528, in factory\r\n    return xla_client.make_c_api_client(\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/jaxlib/xla_client.py"", line 153, in make_c_api_client\r\n    return _xla.get_c_api_client(plugin_name, options, distributed_client)\r\njaxlib._jax.XlaRuntimeError: FAILED_PRECONDITION: No visible GPU devices.\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nTraceback (most recent call last):\r\n  File ""/fast/home/franz.srambical/jafar/train_dynamics.py"", line 157, in <module>\r\n    num_devices = jax.device_count()\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/jax/_src/xla_bridge.py"", line 907, in device_count\r\n    return int(get_backend(backend).device_count())\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/jax/_src/xla_bridge.py"", line 876, in get_backend\r\n    return _get_backend_uncached(platform)\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/jax/_src/xla_bridge.py"", line 855, in _get_backend_uncached\r\n    bs = backends()\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/jax/_src/xla_bridge.py"", line 758, in backends\r\n    raise RuntimeError(err_msg)\r\nRuntimeError: Unable to initialize backend 'cuda': FAILED_PRECONDITION: No visible GPU devices. (you may need to uninstall the failing plugin package, or set JAX_PLATFORMS=cpu to skip this backend.)\r\n",,terminal_output
+155,326144,"TERMINAL",0,0,"srun: error: hai004: task 0: Exited with exit code 1\r\n]0;franz.srambical@hai-login1:~/jafar",,terminal_output
+156,351426,"TERMINAL",0,0,"salloc --gpus=1 --ntasks-per-node=1 --cpus-per-task=1 --mem=100G",,terminal_command
+157,351485,"TERMINAL",0,0,"]633;Csalloc: Granted job allocation 17482\r\n",,terminal_output
+158,351603,"TERMINAL",0,0,"salloc: Waiting for resource configuration\r\n",,terminal_output
+159,352596,"TERMINAL",0,0,"salloc: Nodes hai004 are ready for job\r\n",,terminal_output
+160,352982,"TERMINAL",0,0,"Running inside SLURM, Job ID 17482.\r\n",,terminal_output
+161,353054,"TERMINAL",0,0,"]0;franz.srambical@hai-login1:~/jafar[?2004h[franz.srambical@hai004.haicore.berlin:~/jafar] $ ",,terminal_output
+162,353604,"TERMINAL",0,0,"salloc --gpus=1 --ntasks-per-node=1 --cpus-per-task=1 --mem=100G",,terminal_output
+163,353757,"TERMINAL",0,0,"\rcancel --me",,terminal_output
+164,354722,"TERMINAL",0,0,"queue",,terminal_output
+165,355234,"TERMINAL",0,0,"[1@cancel",,terminal_output
+166,355430,"TERMINAL",0,0,"alloc --gpus=1 --ntasks-per-node=1 --cpus-per-task=1 --mem=100G\r",,terminal_output
+167,355599,"TERMINAL",0,0,"",,terminal_output
+168,355786,"TERMINAL",0,0,"",,terminal_output
+169,356110,"TERMINAL",0,0,"\r(reverse-i-search)`': ",,terminal_output
+170,356454,"TERMINAL",0,0,"d': salloc --gpus=1 --ntasks-per-node=1 --cpus-per-task=1 --mem=100G\ry': bash experiments/dynamics_grain_tok_restore.sh [1@n': bash experiments/dyn",,terminal_output
+171,357016,"TERMINAL",0,0,"\r[25@[franz.srambical@hai004.haicore.berlin:~/jafar] $ bash experiments/dyn",,terminal_output
+172,357334,"TERMINAL",0,0,"\r\n[?2004l\r",,terminal_output
+173,361531,"TERMINAL",0,0,"Running on 1 devices.\r\n",,terminal_output
+174,367551,"TERMINAL",0,0,"Counting all components: ['dynamics', 'lam', 'tokenizer']\r\nParameter counts:\r\n{'dynamics': 26555392, 'lam': 35115232, 'tokenizer': 33750256, 'total': 95420880}\r\n",,terminal_output
+175,369186,"TERMINAL",0,0,"WARNING:absl:Metadata file does not exist: /home/franz.srambical/jafar/checkpoints/causal_dynamics_openai_grain_tok_restore/000290/_CHECKPOINT_METADATA\r\n",,terminal_output
+176,369756,"TERMINAL",0,0,"/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/orbax/checkpoint/_src/serialization/type_handlers.py:1256: UserWarning: Sharding info not provided when restoring. Populating sharding info from sharding file. Please note restoration time will be slightly increased due to reading from file. Note also that this option is unsafe when restoring on a different topology than the checkpoint was saved with.\r\n  warnings.warn(\r\n",,terminal_output
+177,371346,"TERMINAL",0,0,"Starting training from step 0...\r\n",,terminal_output
+178,371680,"TERMINAL",0,0,"2025-08-12 11:05:46.511276: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:467] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered\r\nWARNING: All log messages before absl::InitializeLog() is called are written to STDERR\r\nE0000 00:00:1754989546.525447 1645754 cuda_dnn.cc:8579] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered\r\nE0000 00:00:1754989546.529908 1645754 cuda_blas.cc:1407] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered\r\nW0000 00:00:1754989546.541810 1645754 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.\r\nW0000 00:00:1754989546.541826 1645754 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.\r\nW0000 00:00:1754989546.541829 1645754 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.\r\nW0000 00:00:1754989546.541831 1645754 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.\r\n",,terminal_output
+179,386025,"TERMINAL",0,0,"2025-08-12 11:06:00.890997: W external/xla/xla/service/gpu/autotuning/dot_search_space.cc:200] All configs were filtered out because none of them sufficiently match the hints. Maybe the hints set does not contain a good representative set of valid configs?Working around this by using the full hints set instead.\r\n2025-08-12 11:06:00.891500: W external/xla/xla/service/gpu/autotuning/dot_search_space.cc:200] All configs were filtered out because none of them sufficiently match the hints. Maybe the hints set does not contain a good representative set of valid configs?Working around this by using the full hints set instead.\r\n",,terminal_output
+180,399564,"TERMINAL",0,0,"Step 0, loss: 16.592443466186523\r\n",,terminal_output
+181,422974,"TERMINAL",0,0,"Step 1, loss: 2.159707946702838e-05\r\n",,terminal_output
+182,423923,"TERMINAL",0,0,"Step 2, loss: 4.451984523257124e-07\r\n",,terminal_output
+183,424854,"TERMINAL",0,0,"Step 3, loss: 3.7008923925441195e-08\r\n",,terminal_output
+184,425785,"TERMINAL",0,0,"Step 4, loss: 5.766077837421335e-09\r\n",,terminal_output
+185,426744,"TERMINAL",0,0,"Step 5, loss: 1.2309643304675433e-09\r\n",,terminal_output
+186,427693,"TERMINAL",0,0,"Step 6, loss: 3.239382373454447e-10\r\n",,terminal_output
+187,428617,"TERMINAL",0,0,"Step 7, loss: 8.098456627525508e-11\r\n",,terminal_output
+188,429778,"TERMINAL",0,0,"Step 8, loss: 2.4295369535631828e-11\r\n",,terminal_output
+189,430992,"TERMINAL",0,0,"Step 9, loss: 0.0\r\n",,terminal_output
+190,431720,"TERMINAL",0,0,"Step 10, loss: 0.0\r\n",,terminal_output
+191,432626,"TERMINAL",0,0,"Step 11, loss: 0.0\r\n",,terminal_output
+192,433725,"TERMINAL",0,0,"Step 12, loss: 0.0\r\n",,terminal_output
+193,434594,"TERMINAL",0,0,"Step 13, loss: 0.0\r\n",,terminal_output
+194,435524,"TERMINAL",0,0,"Step 14, loss: 0.0\r\n",,terminal_output
+195,436404,"TERMINAL",0,0,"Step 15, loss: 0.0\r\n",,terminal_output
+196,437372,"TERMINAL",0,0,"Step 16, loss: 0.0\r\n",,terminal_output
+197,438621,"TERMINAL",0,0,"Step 17, loss: 0.0\r\n",,terminal_output
+198,439595,"TERMINAL",0,0,"Step 18, loss: 0.0\r\n",,terminal_output
+199,440543,"TERMINAL",0,0,"Step 19, loss: 0.0\r\n",,terminal_output
+200,441475,"TERMINAL",0,0,"Step 20, loss: 0.0\r\n",,terminal_output
+201,442440,"TERMINAL",0,0,"Step 21, loss: 0.0\r\n",,terminal_output
+202,443402,"TERMINAL",0,0,"Step 22, loss: 0.0\r\n",,terminal_output
+203,444327,"TERMINAL",0,0,"Step 23, loss: 0.0\r\n",,terminal_output
+204,445278,"TERMINAL",0,0,"Step 24, loss: 0.0\r\n",,terminal_output
+205,446503,"TERMINAL",0,0,"Step 25, loss: 0.0\r\n",,terminal_output
+206,447422,"TERMINAL",0,0,"Step 26, loss: 0.0\r\n",,terminal_output
+207,448593,"TERMINAL",0,0,"Step 27, loss: 0.0\r\n",,terminal_output
+208,449449,"TERMINAL",0,0,"Step 28, loss: 0.0\r\n",,terminal_output
+209,450358,"TERMINAL",0,0,"Step 29, loss: 0.0\r\n",,terminal_output
+210,451133,"TERMINAL",0,0,"Step 30, loss: 0.0\r\n",,terminal_output
+211,452268,"TERMINAL",0,0,"Step 31, loss: 0.0\r\n",,terminal_output
+212,453270,"TERMINAL",0,0,"Step 32, loss: 0.0\r\n",,terminal_output
+213,454175,"TERMINAL",0,0,"Step 33, loss: 0.0\r\n",,terminal_output
+214,454943,"TERMINAL",0,0,"Step 34, loss: 0.0\r\n",,terminal_output
+215,455900,"TERMINAL",0,0,"Step 35, loss: 0.0\r\n",,terminal_output
+216,456879,"TERMINAL",0,0,"Step 36, loss: 0.0\r\n",,terminal_output
+217,457848,"TERMINAL",0,0,"Step 37, loss: 0.0\r\n",,terminal_output
+218,458820,"TERMINAL",0,0,"Step 38, loss: 0.0\r\n",,terminal_output
+219,459778,"TERMINAL",0,0,"Step 39, loss: 0.0\r\n",,terminal_output
+220,461030,"TERMINAL",0,0,"Step 40, loss: 0.0\r\n",,terminal_output
+221,462294,"TERMINAL",0,0,"Step 41, loss: 0.0\r\n",,terminal_output
+222,463020,"TERMINAL",0,0,"Step 42, loss: 0.0\r\n",,terminal_output
+223,463899,"TERMINAL",0,0,"Step 43, loss: 0.0\r\n",,terminal_output
+224,464862,"TERMINAL",0,0,"Step 44, loss: 0.0\r\n",,terminal_output
+225,465812,"TERMINAL",0,0,"Step 45, loss: 0.0\r\n",,terminal_output
+226,466857,"TERMINAL",0,0,"Step 46, loss: 0.0\r\n",,terminal_output
+227,467939,"TERMINAL",0,0,"Step 47, loss: 0.0\r\n",,terminal_output
+228,468984,"TERMINAL",0,0,"Step 48, loss: 0.0\r\n",,terminal_output
+229,470108,"TERMINAL",0,0,"Step 49, loss: 0.0\r\n",,terminal_output
+230,470991,"TERMINAL",0,0,"Step 50, loss: 0.0\r\n",,terminal_output
+231,471900,"TERMINAL",0,0,"Step 51, loss: 0.0\r\n",,terminal_output
+232,472770,"TERMINAL",0,0,"Step 52, loss: 0.0\r\n",,terminal_output
+233,473619,"TERMINAL",0,0,"Step 53, loss: 0.0\r\n",,terminal_output
+234,474563,"TERMINAL",0,0,"Step 54, loss: 0.0\r\n",,terminal_output
+235,475450,"TERMINAL",0,0,"Step 55, loss: 0.0\r\n",,terminal_output
+236,476072,"TERMINAL",0,0,"^Csrun: interrupt (one more within 1 sec to abort)\r\nsrun: StepId=17482.0 task 0: running\r\n",,terminal_output
+237,476249,"TERMINAL",0,0,"^Csrun: sending Ctrl-C to StepId=17482.0\r\nsrun: forcing job termination\r\nsrun: Job step aborted: Waiting up to 32 seconds for job step to finish.\r\n[2025-08-12T11:07:31.115] error: *** STEP 17482.0 ON hai004 CANCELLED AT 2025-08-12T11:07:31 DUE to SIGNAL Killed ***\r\n",,terminal_output
+238,476449,"TERMINAL",0,0,"^Csrun: sending Ctrl-C to StepId=17482.0\r\nsrun: job abort in progress\r\n",,terminal_output
+239,476871,"TERMINAL",0,0,"]0;franz.srambical@hai-login1:~/jafar[?2004h[franz.srambical@hai004.haicore.berlin:~/jafar] $ ",,terminal_output
+240,479564,"train_dynamics.py",12324,0,"",python,selection_command
+241,483571,"train_dynamics.py",13259,0,"",python,selection_command
+242,497961,"train_dynamics.py",14606,0,"",python,selection_command
+243,498104,"train_dynamics.py",14982,0,"",python,selection_command
+244,498867,"train_dynamics.py",14947,0,"",python,selection_command
+245,499124,"train_dynamics.py",14946,0,"",python,selection_command
+246,499137,"train_dynamics.py",14920,0,"",python,selection_command
+247,499440,"train_dynamics.py",14877,0,"",python,selection_command
+248,500174,"train_dynamics.py",14893,0,"",python,selection_command
+249,500334,"train_dynamics.py",14896,0,"",python,selection_command
+250,500507,"train_dynamics.py",14901,0,"",python,selection_command
+251,500685,"train_dynamics.py",14904,0,"",python,selection_command
+252,501144,"train_dynamics.py",14904,1,"a",python,selection_command
+253,501214,"train_dynamics.py",14904,4,"args",python,selection_command
+254,501396,"train_dynamics.py",14904,5,"args.",python,selection_command
+255,501619,"train_dynamics.py",14904,14,"args.num_steps",python,selection_command
+256,502744,"train_dynamics.py",14904,14,"",python,content
+257,503329,"train_dynamics.py",14904,0,"5",python,content
+258,503329,"train_dynamics.py",14905,0,"",python,selection_keyboard
+259,503499,"train_dynamics.py",14904,0,"",python,selection_command
+260,506573,"TERMINAL",0,0,"bash experiments/dynamics_grain_tok_restore.sh ",,terminal_output
+261,507768,"TERMINAL",0,0,"\r\n[?2004l\r",,terminal_output
+262,511572,"TERMINAL",0,0,"Running on 1 devices.\r\n",,terminal_output
+263,517628,"TERMINAL",0,0,"Counting all components: ['dynamics', 'lam', 'tokenizer']\r\nParameter counts:\r\n{'dynamics': 26555392, 'lam': 35115232, 'tokenizer': 33750256, 'total': 95420880}\r\n",,terminal_output
+264,519246,"TERMINAL",0,0,"WARNING:absl:Metadata file does not exist: /home/franz.srambical/jafar/checkpoints/causal_dynamics_openai_grain_tok_restore/000290/_CHECKPOINT_METADATA\r\n",,terminal_output
+265,519812,"TERMINAL",0,0,"/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/orbax/checkpoint/_src/serialization/type_handlers.py:1256: UserWarning: Sharding info not provided when restoring. Populating sharding info from sharding file. Please note restoration time will be slightly increased due to reading from file. Note also that this option is unsafe when restoring on a different topology than the checkpoint was saved with.\r\n  warnings.warn(\r\n",,terminal_output
+266,521564,"TERMINAL",0,0,"Starting training from step 0...\r\n",,terminal_output
+267,521909,"TERMINAL",0,0,"2025-08-12 11:08:16.585319: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:467] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered\r\nWARNING: All log messages before absl::InitializeLog() is called are written to STDERR\r\nE0000 00:00:1754989696.599650 1744785 cuda_dnn.cc:8579] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered\r\nE0000 00:00:1754989696.604072 1744785 cuda_blas.cc:1407] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered\r\nW0000 00:00:1754989696.615729 1744785 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.\r\nW0000 00:00:1754989696.615747 1744785 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.\r\nW0000 00:00:1754989696.615753 1744785 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.\r\nW0000 00:00:1754989696.615755 1744785 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once.\r\n",,terminal_output
+268,534751,"TERMINAL",0,0,"2025-08-12 11:08:29.626720: W external/xla/xla/service/gpu/autotuning/dot_search_space.cc:200] All configs were filtered out because none of them sufficiently match the hints. Maybe the hints set does not contain a good representative set of valid configs?Working around this by using the full hints set instead.\r\n2025-08-12 11:08:29.627236: W external/xla/xla/service/gpu/autotuning/dot_search_space.cc:200] All configs were filtered out because none of them sufficiently match the hints. Maybe the hints set does not contain a good representative set of valid configs?Working around this by using the full hints set instead.\r\n",,terminal_output

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-5844baec-254e-4a51-8bbd-690c56cee10a1759049586350-2025_09_28-10.53.34.166/source.csv

@@ -0,0 +1,5 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"data/.venv/lib/python3.10/site-packages/procgen/env.py",0,0,"import os\nimport random\nfrom typing import Sequence, Optional, List\n\nimport gym3\nfrom gym3.libenv import CEnv\nimport numpy as np\nfrom .builder import build\n\nSCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))\n\nMAX_STATE_SIZE = 2 ** 20\n\nENV_NAMES = [\n    ""bigfish"",\n    ""bossfight"",\n    ""caveflyer"",\n    ""chaser"",\n    ""climber"",\n    ""coinrun"",\n    ""dodgeball"",\n    ""fruitbot"",\n    ""heist"",\n    ""jumper"",\n    ""leaper"",\n    ""maze"",\n    ""miner"",\n    ""ninja"",\n    ""plunder"",\n    ""starpilot"",\n]\n\nEXPLORATION_LEVEL_SEEDS = {\n    ""coinrun"": 1949448038,\n    ""caveflyer"": 1259048185,\n    ""leaper"": 1318677581,\n    ""jumper"": 1434825276,\n    ""maze"": 158988835,\n    ""heist"": 876640971,\n    ""climber"": 1561126160,\n    ""ninja"": 1123500215,\n}\n\n# should match DistributionMode in game.h, except for 'exploration' which is handled by Python\nDISTRIBUTION_MODE_DICT = {\n    ""easy"": 0,\n    ""hard"": 1,\n    ""extreme"": 2,\n    ""memory"": 10,\n    ""exploration"": 20,\n}\n\n\ndef create_random_seed():\n    rand_seed = random.SystemRandom().randint(0, 2 ** 31 - 1)\n    try:\n        # force MPI processes to definitely choose different random seeds\n        from mpi4py import MPI\n\n        rand_seed = rand_seed - (rand_seed % MPI.COMM_WORLD.size) + MPI.COMM_WORLD.rank\n    except ModuleNotFoundError:\n        pass\n    return rand_seed\n\n\nclass BaseProcgenEnv(CEnv):\n    """"""\n    Base procedurally generated environment\n    """"""\n\n    def __init__(\n        self,\n        num,\n        env_name,\n        options,\n        debug=False,\n        rand_seed=None,\n        num_levels=0,\n        start_level=0,\n        use_sequential_levels=False,\n        debug_mode=0,\n        resource_root=None,\n        num_threads=4,\n        render_mode=None,\n    ):\n        if resource_root is None:\n            resource_root = os.path.join(SCRIPT_DIR, ""data"", ""assets"") + os.sep\n            assert os.path.exists(resource_root)\n\n        lib_dir = os.path.join(SCRIPT_DIR, ""data"", ""prebuilt"")\n        if os.path.exists(lib_dir):\n            assert any([os.path.exists(os.path.join(lib_dir, name)) for name in [""libenv.so"", ""libenv.dylib"", ""env.dll""]]), ""package is installed, but the prebuilt environment library is missing""\n            assert not debug, ""debug has no effect for pre-compiled library""\n        else:\n            # only compile if we don't find a pre-built binary\n            lib_dir = build(debug=debug)\n        \n        self.combos = self.get_combos()\n\n        if render_mode is None:\n            render_human = False\n        elif render_mode == ""rgb_array"":\n            render_human = True\n        else:\n            raise Exception(f""invalid render mode {render_mode}"")\n\n        if rand_seed is None:\n            rand_seed = create_random_seed()\n\n        options.update(\n            {\n                ""env_name"": env_name,\n                ""num_levels"": num_levels,\n                ""start_level"": start_level,\n                ""num_actions"": len(self.combos),\n                ""use_sequential_levels"": bool(use_sequential_levels),\n                ""debug_mode"": debug_mode,\n                ""rand_seed"": rand_seed,\n                ""num_threads"": num_threads,\n                ""render_human"": render_human,\n                # these will only be used the first time an environment is created in a process\n                ""resource_root"": resource_root,\n            }\n        )\n\n        self.options = options\n\n        super().__init__(\n            lib_dir=lib_dir,\n            num=num,\n            options=options,\n            c_func_defs=[\n                ""int get_state(libenv_env *, int, char *, int);"",\n                ""void set_state(libenv_env *, int, char *, int);"",\n            ],\n        )\n        # don't use the dict space for actions\n        self.ac_space = self.ac_space[""action""]\n\n    def get_state(self):\n        length = MAX_STATE_SIZE\n        buf = self._ffi.new(f""char[{length}]"")\n        result = []\n        for env_idx in range(self.num):\n            n = self.call_c_func(""get_state"", env_idx, buf, length)\n            result.append(bytes(self._ffi.buffer(buf, n)))\n        return result\n\n    def set_state(self, states):\n        assert len(states) == self.num\n        for env_idx in range(self.num):\n            state = states[env_idx]\n            self.call_c_func(""set_state"", env_idx, state, len(state))\n\n    def get_combos(self):\n        return [\n            (""LEFT"", ""DOWN""),\n            (""LEFT"",),\n            (""LEFT"", ""UP""),\n            (""DOWN"",),\n            (),\n            (""UP"",),\n            (""RIGHT"", ""DOWN""),\n            (""RIGHT"",),\n            (""RIGHT"", ""UP""),\n            (""D"",),\n            (""A"",),\n            (""W"",),\n            (""S"",),\n            (""Q"",),\n            (""E"",),\n        ]\n\n    def keys_to_act(self, keys_list: Sequence[Sequence[str]]) -> List[Optional[np.ndarray]]:\n        """"""\n        Convert list of keys being pressed to actions, used in interactive mode\n        """"""\n        result = []\n        for keys in keys_list:\n            action = None\n            max_len = -1\n            for i, combo in enumerate(self.get_combos()):\n                pressed = True\n                for key in combo:\n                    if key not in keys:\n                        pressed = False\n\n                if pressed and (max_len < len(combo)):\n                    action = i\n                    max_len = len(combo)\n\n            if action is not None:\n                action = np.array([action])\n            result.append(action)\n        return result\n\n    def act(self, ac):\n        # tensorflow may return int64 actions (https://github.com/openai/gym/blob/master/gym/spaces/discrete.py#L13)\n        # so always cast actions to int32\n        return super().act({""action"": ac.astype(np.int32)})\n\n\nclass ProcgenGym3Env(BaseProcgenEnv):\n    """"""\n    gym3 interface for Procgen\n    """"""\n    def __init__(\n        self,\n        num,\n        env_name,\n        center_agent=True,\n        use_backgrounds=True,\n        use_monochrome_assets=False,\n        restrict_themes=False,\n        use_generated_assets=False,\n        paint_vel_info=False,\n        distribution_mode=""hard"",\n        **kwargs,\n    ):\n        assert (\n            distribution_mode in DISTRIBUTION_MODE_DICT\n        ), f'""{distribution_mode}"" is not a valid distribution mode.'\n\n        if distribution_mode == ""exploration"":\n            assert (\n                env_name in EXPLORATION_LEVEL_SEEDS\n            ), f""{env_name} does not support exploration mode""\n\n            distribution_mode = DISTRIBUTION_MODE_DICT[""hard""]\n            assert ""num_levels"" not in kwargs, ""exploration mode overrides num_levels""\n            kwargs[""num_levels""] = 1\n            assert ""start_level"" not in kwargs, ""exploration mode overrides start_level""\n            kwargs[""start_level""] = EXPLORATION_LEVEL_SEEDS[env_name]\n        else:\n            distribution_mode = DISTRIBUTION_MODE_DICT[distribution_mode]\n\n        options = {\n                ""center_agent"": bool(center_agent),\n                ""use_generated_assets"": bool(use_generated_assets),\n                ""use_monochrome_assets"": bool(use_monochrome_assets),\n                ""restrict_themes"": bool(restrict_themes),\n                ""use_backgrounds"": bool(use_backgrounds),\n                ""paint_vel_info"": bool(paint_vel_info),\n                ""distribution_mode"": distribution_mode,\n            }\n        super().__init__(num, env_name, options, **kwargs)\n        \n        \nclass ToBaselinesVecEnv(gym3.ToBaselinesVecEnv):\n    metadata = {\n        'render.modes': ['human', 'rgb_array'],\n        'video.frames_per_second' : 15\n    }\n    def render(self, mode=""human""):\n        info = self.env.get_info()[0]\n        _, ob, _ = self.env.observe()\n        if mode == ""rgb_array"":\n            if ""rgb"" in info:\n                return info[""rgb""]\n            else:\n                return ob['rgb'][0]        \n\n\ndef ProcgenEnv(num_envs, env_name, **kwargs):\n    return ToBaselinesVecEnv(ProcgenGym3Env(num=num_envs, env_name=env_name, **kwargs))\n",python,tab
+2,430,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"10:53:34 AM [info] Activating crowd-code\n10:53:34 AM [info] Recording started\n10:53:34 AM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,510,"extension-output-pdoom-org.crowd-code-#1-crowd-code",153,0,"10:53:34 AM [info] Git repository found\n10:53:34 AM [info] Git provider initialized successfully\n10:53:34 AM [info] Initial git state: [object Object]\n",Log,content
+4,98529,"TERMINAL",0,0,"",,terminal_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-66c4f35a-3491-46c3-8cc2-2d58bb70fa771767626653190-2026_01_05-16.24.25.952/source.csv

@@ -0,0 +1,346 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"package.json",0,0,"{\n  ""name"": ""crowd-pilot"",\n  ""displayName"": ""crowd-pilot-extension"",\n  ""description"": ""Teaching language models to code like humans."",\n  ""publisher"": ""p-doom"",\n  ""version"": ""0.0.1"",\n  ""repository"": {\n    ""type"": ""git"",\n    ""url"": ""https://github.com/p-doom/crowd-pilot-extension""\n  },\n  ""engines"": {\n    ""vscode"": ""^1.99.3""\n  },\n  ""categories"": [\n    ""Other""\n  ],\n  ""activationEvents"": [\n    ""onStartupFinished""\n  ],\n  ""main"": ""./out/extension.js"",\n  ""contributes"": {\n    ""commands"": [\n      {\n        ""command"": ""crowd-pilot.toggleSuggestions"",\n        ""title"": ""crowd-pilot: Toggle Tab Suggestions""\n      },\n      {\n        ""command"": ""crowd-pilot.hideUi"",\n        ""title"": ""crowd-pilot: Hide Preview""\n      },\n      {\n        ""command"": ""crowd-pilot.sglangTest"",\n        ""title"": ""crowd-pilot: Test SGLang""\n      },\n      {\n        ""command"": ""crowd-pilot.modelRun"",\n        ""title"": ""crowd-pilot: Model Plan & Run""\n      },\n      {\n        ""command"": ""crowd-pilot.clearContext"",\n        ""title"": ""crowd-pilot: Clear Context""\n      },\n      {\n        ""command"": ""crowd-pilot.openPreferenceLog"",\n        ""title"": ""crowd-pilot: Open Preference Log""\n      },\n      {\n        ""command"": ""crowd-pilot.showPendingAction"",\n        ""title"": ""crowd-pilot: Show Pending Suggestion""\n      }\n    ],\n    ""configuration"": {\n      ""title"": ""crowd-pilot"",\n      ""properties"": {\n        ""crowd-pilot.hostname"": {\n          ""type"": ""string"",\n          ""default"": ""hai002"",\n          ""description"": ""Hostname of the SGLang server""\n        },\n        ""crowd-pilot.port"": {\n          ""type"": ""number"",\n          ""default"": 30000,\n          ""description"": ""Port of the SGLang server""\n        },\n        ""crowd-pilot.basePath"": {\n          ""type"": ""string"",\n          ""default"": ""/v1/chat/completions"",\n          ""description"": ""Base path for the SGLang API endpoint""\n        },\n        ""crowd-pilot.modelName"": {\n          ""type"": ""string"",\n          ""default"": ""qwen/qwen3-8b"",\n          ""description"": ""Model name to use for completions""\n        },\n        ""crowd-pilot.minAvgLogprob"": {\n          ""type"": ""number"",\n          ""default"": -1.0,\n          ""description"": ""Minimum average log-probability per token for displaying suggestions. Higher values (closer to 0) require more confidence. -1.0 ≈ perplexity 2.7""\n        },\n        ""crowd-pilot.maxContextTokens"": {\n          ""type"": ""number"",\n          ""default"": 120000,\n          ""description"": ""Context length (in tokens). Older messages are truncated to fit. Set below your model's limit to leave room for the response.""\n        },\n        ""crowd-pilot.enablePreferenceLogging"": {\n          ""type"": ""boolean"",\n          ""default"": true,\n          ""description"": ""Enable logging of accept/reject data for reward model training and RLHF/DPO""\n        },\n        ""crowd-pilot.preferenceLogPath"": {\n          ""type"": ""string"",\n          ""default"": """",\n          ""description"": ""Custom path for the preference log file (JSONL format). If empty, uses workspace/.crowd-pilot-preferences.jsonl""\n        },\n        ""crowd-pilot.viewportRadius"": {\n          ""type"": ""number"",\n          ""default"": 10,\n          ""description"": ""Number of lines above/below cursor to include in file viewports""\n        }\n      }\n    },\n    ""keybindings"": [\n      {\n        ""command"": ""crowd-pilot.modelRun"",\n        ""key"": ""tab"",\n        ""mac"": ""tab"",\n        ""when"": ""editorTextFocus && crowdPilot.uiVisible""\n      },\n      {\n        ""command"": ""crowd-pilot.modelRun"",\n        ""key"": ""tab"",\n        ""mac"": ""tab"",\n        ""when"": ""inQuickOpen && crowdPilot.uiVisible""\n      },\n      {\n        ""command"": ""crowd-pilot.hideUi"",\n        ""key"": ""escape"",\n        ""mac"": ""escape"",\n        ""when"": ""crowdPilot.uiVisible""\n      },\n      {\n        ""command"": ""crowd-pilot.showPendingAction"",\n        ""key"": ""ctrl+shift+space"",\n        ""mac"": ""cmd+shift+space"",\n        ""when"": ""terminalFocus && crowdPilot.hasPendingAction""\n      }\n    ]\n  },\n  ""scripts"": {\n    ""vscode:prepublish"": ""npm run compile"",\n    ""compile"": ""tsc -p ./"",\n    ""watch"": ""tsc -watch -p ./"",\n    ""pretest"": ""npm run compile && npm run lint"",\n    ""lint"": ""eslint src"",\n    ""test"": ""vscode-test"",\n    ""clean"": ""rm -rf out *.tgz"",\n    ""clean:all"": ""rm -rf out *.tgz node_modules package-lock.json"",\n    ""rebuild-serializer"": ""cd crowd-pilot-serializer/crates/napi && npm install && rm -f index.d.ts index.js && npm run build && npm pack && mv *.tgz ../../../ && cd ../../.. && rm -rf node_modules/@crowd-pilot && npm install""\n  },\n  ""dependencies"": {\n    ""@crowd-pilot/serializer"": ""file:./crowd-pilot-serializer-0.1.0.tgz""\n  },\n  ""devDependencies"": {\n    ""@types/vscode"": ""^1.99.3"",\n    ""@types/mocha"": ""^10.0.10"",\n    ""@types/node"": ""22.x"",\n    ""@typescript-eslint/eslint-plugin"": ""^8.45.0"",\n    ""@typescript-eslint/parser"": ""^8.45.0"",\n    ""eslint"": ""^9.36.0"",\n    ""typescript"": ""^5.9.3"",\n    ""@vscode/test-cli"": ""^0.0.11"",\n    ""@vscode/test-electron"": ""^2.5.2""\n  }\n}\n",json,tab
+2,494,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"4:24:25 PM [info] Activating crowd-code\n4:24:25 PM [info] Recording started\n4:24:25 PM [info] Initializing git provider using file system watchers...\n4:24:26 PM [info] Git repository found\n4:24:26 PM [info] Git provider initialized successfully\n4:24:26 PM [info] Initial git state: [object Object]\n",Log,tab
+3,1286,"package.json",0,0,"",json,tab
+4,1671,"package.json",135,0,"",json,selection_command
+5,1740,"package.json",160,0,"",json,selection_command
+6,6582,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+7,9846,"TERMINAL",0,0,"",,terminal_focus
+8,9847,"package.json",0,0,"",json,tab
+9,11571,"TERMINAL",0,0,"squeue",,terminal_command
+10,11572,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  39368 nick.lecht  standard     1   128  R 2026-01-05T15:24:15 2026-01-05T15:24:15  1:00:22   12:00:00 hai006\r\n  39367 nick.lecht  standard     1   128  R 2026-01-05T15:22:56 2026-01-05T15:22:56  1:01:41   12:00:00 hai002\r\n  39351 nishant.ku  standard     3   624  R 2026-01-05T10:44:33 2026-01-05T10:44:33  5:40:04 1-00:00:00 hai[003-005]\r\n]0;franz.srambical@hai-login2:~/crowd-pilot-extension",,terminal_output
+11,17084,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",0,0,"#!/usr/bin/env bash\n\n#SBATCH --nodes=1\n#SBATCH --ntasks-per-node=1\n#SBATCH --time=24:00:00\n#SBATCH --cpus-per-task=8\n#SBATCH --gres=gpu:4\n#SBATCH --output=/fast/project/HFMI_SynergyUnit/tab_model/logs/franz/%x_%j.log\n#SBATCH --error=/fast/project/HFMI_SynergyUnit/tab_model/logs/franz/%x_%j.log\n#SBATCH --job-name=crowd_pilot_sglang\n#SBATCH --mem=400GB\n#SBATCH --qos=normal\n\nexport HF_HOME=/fast/project/HFMI_SynergyUnit/tab_model/franz/hf_home/\n\nsource /home/franz.srambical/crowd-pilot-serializer-legacy/.venv/bin/activate\nmodule load CUDA/12.8\n\nmodel_path=""zai-org/GLM-4.5-Air""\npython3 -m sglang.launch_server --model-path $model_path --host 0.0.0.0 --log-requests \\n  --tp-size 4 \\n  --tool-call-parser glm45  \\n  --reasoning-parser glm45 \\n  --speculative-algorithm EAGLE \\n  --speculative-num-steps 3 \\n  --speculative-eagle-topk 1 \\n  --speculative-num-draft-tokens 4 \\n  --mem-fraction-static 0.9",shellscript,tab
+12,24418,"TERMINAL",0,0,"sbatch /home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",,terminal_command
+13,24419,"TERMINAL",0,0,"]633;CSubmitted batch job 39373\r\n]0;franz.srambical@hai-login2:~/crowd-pilot-extension",,terminal_output
+14,504474,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",157,0,"",shellscript,selection_command
+15,504626,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",236,0,"",shellscript,selection_command
+16,512553,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",314,0,"",shellscript,selection_command
+17,512595,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",236,0,"",shellscript,selection_command
+18,515061,"TERMINAL",0,0,"squeue -u franz.srambical",,terminal_command
+19,515092,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  39373 franz.sram  standard     1    16  R 2026-01-05T16:24:50 2026-01-05T16:24:50     8:10 1-00:00:00 hai001\r\n]0;franz.srambical@hai-login2:~/crowd-pilot-extension",,terminal_output
+20,526109,"src/preview/quickPick.ts",0,0,"import * as vscode from 'vscode';\nimport { Action, truncate } from './types';\n\n/**\n * Result of the quick pick interaction.\n */\nexport type QuickPickResult = 'accept' | 'dismiss' | null;\n\n/**\n * Show a quick pick modal for the pending action.\n * Used when terminal is focused and decorations can't be shown.\n */\nexport async function showPendingActionQuickPick(action: Action): Promise<QuickPickResult> {\n    const detail = formatActionDetail(action);\n    \n    const items: vscode.QuickPickItem[] = [\n        { \n            label: '$(check) Accept', \n            description: 'Execute this action',\n            detail: detail\n        },\n        { \n            label: '$(x) Dismiss', \n            description: 'Cancel this suggestion'\n        },\n    ];\n\n    const result = await vscode.window.showQuickPick(items, {\n        title: 'Pending Suggestion',\n        placeHolder: getActionSummary(action),\n        ignoreFocusOut: false,\n    });\n\n    if (result?.label.includes('Accept')) {\n        return 'accept';\n    }\n    if (result?.label.includes('Dismiss')) {\n        return 'dismiss';\n    }\n    return null;\n}\n\n/**\n * Get a short summary of the action for the quick pick placeholder.\n */\nfunction getActionSummary(action: Action): string {\n    switch (action.kind) {\n        case 'terminalSendText':\n            return `Run terminal command`;\n        case 'openFile':\n            const fileName = action.filePath.split(/[/\\]/).pop() || action.filePath;\n            return `Open file: ${fileName}`;\n        case 'setSelections':\n            return `Move cursor to line ${action.selections[0].start[0] + 1}`;\n        case 'editInsert':\n            return 'Insert text';\n        case 'editReplace':\n            return 'Replace text';\n        case 'editDelete':\n            return `Delete lines ${action.range.start[0] + 1}–${action.range.end[0] + 1}`;\n        case 'terminalShow':\n            return 'Show terminal';\n        case 'showTextDocument':\n            return 'Show document';\n        default:\n            return 'Execute action';\n    }\n}\n\n/**\n * Format the full action detail for display in quick pick.\n */\nfunction formatActionDetail(action: Action): string {\n    switch (action.kind) {\n        case 'terminalSendText':\n            return action.text;\n        case 'openFile':\n            if (action.selections?.[0]) {\n                const line = action.selections[0].start[0] + 1;\n                return `${action.filePath}:${line}`;\n            }\n            return action.filePath;\n        case 'setSelections':\n            const sel = action.selections[0];\n            return `Line ${sel.start[0] + 1}, Column ${sel.start[1] + 1}`;\n        case 'editInsert':\n            return truncate(action.text, 200);\n        case 'editReplace':\n            return truncate(action.text, 200);\n        case 'editDelete':\n            return `Lines ${action.range.start[0] + 1} to ${action.range.end[0] + 1}`;\n        default:\n            return '';\n    }\n}\n\n\n\n\n",typescript,tab
+21,528400,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",0,0,"",shellscript,tab
+22,628926,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",314,0,"",shellscript,selection_command
+23,740642,"Untitled-1",0,0,"",plaintext,tab
+24,742220,"Untitled-1",0,0,"d",plaintext,content
+25,742220,"Untitled-1",1,0,"",plaintext,selection_keyboard
+26,742317,"Untitled-1",1,0,"e",plaintext,content
+27,742317,"Untitled-1",2,0,"",plaintext,selection_keyboard
+28,742503,"Untitled-1",2,0,"f",plaintext,content
+29,742503,"Untitled-1",3,0,"",plaintext,selection_keyboard
+30,742530,"Untitled-1",3,0," ",plaintext,content
+31,742530,"Untitled-1",4,0,"",plaintext,selection_keyboard
+32,744019,"Untitled-1",4,0,"h",plaintext,content
+33,744020,"Untitled-1",5,0,"",plaintext,selection_keyboard
+34,744076,"Untitled-1",5,0,"e",plaintext,content
+35,744076,"Untitled-1",6,0,"",plaintext,selection_keyboard
+36,744237,"Untitled-1",6,0,"l",plaintext,content
+37,744238,"Untitled-1",7,0,"",plaintext,selection_keyboard
+38,744392,"Untitled-1",7,0,"l",plaintext,content
+39,744393,"Untitled-1",8,0,"",plaintext,selection_keyboard
+40,744612,"Untitled-1",7,1,"",plaintext,content
+41,744771,"Untitled-1",6,1,"",plaintext,content
+42,745025,"Untitled-1",5,1,"",plaintext,content
+43,745401,"Untitled-1",4,1,"",plaintext,content
+44,745535,"Untitled-1",4,0,"f",plaintext,content
+45,745536,"Untitled-1",5,0,"",plaintext,selection_keyboard
+46,745545,"Untitled-1",5,0,"i",plaintext,content
+47,745546,"Untitled-1",6,0,"",plaintext,selection_keyboard
+48,746182,"Untitled-1",6,0,"b",plaintext,content
+49,746182,"Untitled-1",7,0,"",plaintext,selection_keyboard
+50,746304,"Untitled-1",7,0,"o",plaintext,content
+51,746305,"Untitled-1",8,0,"",plaintext,selection_keyboard
+52,746371,"Untitled-1",8,0,"n",plaintext,content
+53,746371,"Untitled-1",9,0,"",plaintext,selection_keyboard
+54,747363,"Untitled-1",9,0,"n",plaintext,content
+55,747363,"Untitled-1",10,0,"",plaintext,selection_keyboard
+56,747521,"Untitled-1",10,0,"a",plaintext,content
+57,747522,"Untitled-1",11,0,"",plaintext,selection_keyboard
+58,747543,"Untitled-1",11,0,"c",plaintext,content
+59,747544,"Untitled-1",12,0,"",plaintext,selection_keyboard
+60,747732,"Untitled-1",12,0,"c",plaintext,content
+61,747733,"Untitled-1",13,0,"",plaintext,selection_keyboard
+62,747806,"Untitled-1",13,0,"i",plaintext,content
+63,747807,"Untitled-1",14,0,"",plaintext,selection_keyboard
+64,748402,"Untitled-1",13,1,"",plaintext,content
+65,748574,"Untitled-1",12,1,"",plaintext,content
+66,749393,"Untitled-1",11,1,"",plaintext,content
+67,750229,"Untitled-1",10,1,"",plaintext,content
+68,750399,"Untitled-1",9,1,"",plaintext,content
+69,750722,"Untitled-1",9,0,"a",plaintext,content
+70,750723,"Untitled-1",10,0,"",plaintext,selection_keyboard
+71,750817,"Untitled-1",10,0,"c",plaintext,content
+72,750817,"Untitled-1",11,0,"",plaintext,selection_keyboard
+73,751105,"Untitled-1",11,0,"c",plaintext,content
+74,751105,"Untitled-1",12,0,"",plaintext,selection_keyboard
+75,751398,"Untitled-1",12,0,"i",plaintext,content
+76,751398,"Untitled-1",13,0,"",plaintext,selection_keyboard
+77,752539,"Untitled-1",13,0,"()",plaintext,content
+78,752539,"Untitled-1",14,0,"",plaintext,selection_keyboard
+79,755496,"Untitled-1",14,0,"n",plaintext,content
+80,755497,"Untitled-1",15,0,"",plaintext,selection_keyboard
+81,756910,"Untitled-1",0,16,"def fibonacci(n):\n",plaintext,content
+82,758649,"Untitled-1",18,0,"    if n <= 1:\n",plaintext,content
+83,769558,"Untitled-1",32,0,"",plaintext,selection_command
+84,769688,"Untitled-1",33,0,"",plaintext,selection_command
+85,771165,"Untitled-1",33,0,"        return n\n",plaintext,content
+86,772075,"Untitled-1",50,0,"    else:\n",plaintext,content
+87,772870,"Untitled-1",60,0,"        return fibonacci(n-1) + fibonacci(n-2)\n",plaintext,content
+88,778939,"Untitled-1",106,1,"",plaintext,content
+89,779077,"Untitled-1",60,46,"",plaintext,content
+90,781042,"Untitled-1",59,1,"",plaintext,content
+91,781188,"Untitled-1",50,9,"",plaintext,content
+92,783730,"Untitled-1",50,0,"    return fibonacci(n-1) + fibonacci(n-2)\n",plaintext,content
+93,785151,"Untitled-1",50,0,"",plaintext,selection_command
+94,790667,"Untitled-1",93,0,"",plaintext,selection_command
+95,791120,"Untitled-1",92,1,"",plaintext,content
+96,791311,"Untitled-1",50,42,"",plaintext,content
+97,792534,"Untitled-1",50,0,"e",plaintext,content
+98,792535,"Untitled-1",51,0,"",plaintext,selection_keyboard
+99,792680,"Untitled-1",51,0,"l",plaintext,content
+100,792680,"Untitled-1",52,0,"",plaintext,selection_keyboard
+101,792812,"Untitled-1",52,0,"s",plaintext,content
+102,792812,"Untitled-1",53,0,"",plaintext,selection_keyboard
+103,792812,"Untitled-1",53,0,"e",plaintext,content
+104,792812,"Untitled-1",54,0,"",plaintext,selection_keyboard
+105,792964,"Untitled-1",54,0," ",plaintext,content
+106,792964,"Untitled-1",55,0,"",plaintext,selection_keyboard
+107,794152,"Untitled-1",50,5,"else:\n",plaintext,content
+108,795316,"Untitled-1",56,0,"    return fibonacci(n-1) + fibonacci(n-2)\n",plaintext,content
+109,796738,"Untitled-1",56,0,"",plaintext,selection_command
+110,796867,"Untitled-1",50,0,"",plaintext,selection_command
+111,799877,"Untitled-1",50,0,"    ",plaintext,content
+112,801158,"Untitled-1",60,43,"        return fibonacci(n-1) + fibonacci(n-2)\n",plaintext,content
+113,804541,"Untitled-1",53,0,"",plaintext,selection_command
+114,813358,"Untitled-1",59,0,"",plaintext,selection_command
+115,813706,"Untitled-1",50,9,"",plaintext,content
+116,815471,"Untitled-1",33,0,"",plaintext,selection_command
+117,815471,"Untitled-1",50,0,"",plaintext,selection_command
+118,816122,"Untitled-1",33,0,"",plaintext,selection_command
+119,816389,"Untitled-1",50,0,"",plaintext,selection_command
+120,818452,"Untitled-1",33,0,"",plaintext,selection_command
+121,818606,"Untitled-1",50,0,"",plaintext,selection_command
+122,820808,"Untitled-1",50,0," ",plaintext,content
+123,820809,"Untitled-1",51,0,"",plaintext,selection_keyboard
+124,820809,"Untitled-1",50,1,"",plaintext,content
+125,823085,"Untitled-1",50,0,"    ",plaintext,content
+126,824677,"Untitled-1",54,0,"e",plaintext,content
+127,824677,"Untitled-1",55,0,"",plaintext,selection_keyboard
+128,826239,"Untitled-1",50,6,"    el\n",plaintext,content
+129,827425,"Untitled-1",50,7,"    else\n",plaintext,content
+130,829389,"Untitled-1",50,9,"    else:\n",plaintext,content
+131,831714,"Untitled-1",54,0,"",plaintext,selection_command
+132,832969,"Untitled-1",50,9,"    ",plaintext,content
+133,833776,"Untitled-1",50,4,"",plaintext,content
+134,836434,"Untitled-1",50,1,"",plaintext,content
+135,838924,"Untitled-1",33,0,"",plaintext,selection_command
+136,839099,"Untitled-1",49,0,"\n        ",plaintext,content
+137,839728,"Untitled-1",54,4,"",plaintext,content
+138,840966,"Untitled-1",50,4,"",plaintext,content
+139,842175,"Untitled-1",33,0,"",plaintext,selection_command
+140,842260,"Untitled-1",50,0,"",plaintext,selection_command
+141,845390,"Untitled-1",33,0,"",plaintext,selection_command
+142,845503,"Untitled-1",18,0,"",plaintext,selection_command
+143,845804,"Untitled-1",0,0,"",plaintext,selection_command
+144,846277,"Untitled-1",0,17,"def fibonacci(n):",plaintext,selection_command
+145,846507,"Untitled-1",0,98,"def fibonacci(n):\n    if n <= 1:\n        return n\n\n        return fibonacci(n-1) + fibonacci(n-2)\n",plaintext,selection_command
+146,846844,"Untitled-1",0,98,"",plaintext,content
+147,847836,"Untitled-1",0,0,"d",plaintext,content
+148,847836,"Untitled-1",1,0,"",plaintext,selection_keyboard
+149,847936,"Untitled-1",1,0,"e",plaintext,content
+150,847936,"Untitled-1",2,0,"",plaintext,selection_keyboard
+151,848083,"Untitled-1",2,0,"f",plaintext,content
+152,848083,"Untitled-1",3,0,"",plaintext,selection_keyboard
+153,848146,"Untitled-1",3,0," ",plaintext,content
+154,848147,"Untitled-1",4,0,"",plaintext,selection_keyboard
+155,849555,"Untitled-1",0,4,"def fib\n",plaintext,content
+156,851642,"Untitled-1",0,8,"def fib(\n",plaintext,content
+157,852704,"Untitled-1",0,9,"def fib(n\n",plaintext,content
+158,853716,"Untitled-1",0,10,"def fib(n):\n",plaintext,content
+159,855945,"Untitled-1",12,0,"    if n <= 1:\n",plaintext,content
+160,857115,"Untitled-1",27,0,"        return n\n",plaintext,content
+161,879384,"Untitled-1",27,0,"",plaintext,selection_command
+162,879492,"Untitled-1",12,0,"",plaintext,selection_command
+163,879633,"Untitled-1",0,0,"",plaintext,selection_command
+164,880404,"Untitled-1",0,1,"d",plaintext,selection_command
+165,880881,"Untitled-1",0,13,"def fib(n):\n ",plaintext,selection_command
+166,880882,"Untitled-1",0,28,"def fib(n):\n    if n <= 1:\n ",plaintext,selection_command
+167,881093,"Untitled-1",0,44,"def fib(n):\n    if n <= 1:\n        return n\n",plaintext,selection_command
+168,881292,"Untitled-1",0,44,"",plaintext,content
+169,913969,"Untitled-1",0,0,"d",plaintext,content
+170,913970,"Untitled-1",1,0,"",plaintext,selection_keyboard
+171,914060,"Untitled-1",1,0,"e",plaintext,content
+172,914060,"Untitled-1",2,0,"",plaintext,selection_keyboard
+173,914272,"Untitled-1",2,0,"f",plaintext,content
+174,914272,"Untitled-1",3,0,"",plaintext,selection_keyboard
+175,914297,"Untitled-1",3,0," ",plaintext,content
+176,914297,"Untitled-1",4,0,"",plaintext,selection_keyboard
+177,914834,"Untitled-1",4,0,"f",plaintext,content
+178,914834,"Untitled-1",5,0,"",plaintext,selection_keyboard
+179,914871,"Untitled-1",5,0,"i",plaintext,content
+180,914871,"Untitled-1",6,0,"",plaintext,selection_keyboard
+181,915116,"Untitled-1",6,0,"b",plaintext,content
+182,915117,"Untitled-1",7,0,"",plaintext,selection_keyboard
+183,916183,"Untitled-1",0,7,"def fib(\n",plaintext,content
+184,917323,"Untitled-1",0,9,"def fib(n\n",plaintext,content
+185,918231,"Untitled-1",0,10,"def fib(n):\n",plaintext,content
+186,919284,"Untitled-1",12,0,"    if n <= 1:\n",plaintext,content
+187,920240,"Untitled-1",27,0,"        return n\n",plaintext,content
+188,921041,"Untitled-1",44,0,"    return fib(n-1) + fib(n-2)\n",plaintext,content
+189,922849,"Untitled-1",44,0,"",plaintext,selection_command
+190,923035,"Untitled-1",27,0,"",plaintext,selection_command
+191,925027,"Untitled-1",43,0,"\n        ",plaintext,content
+192,925620,"Untitled-1",48,4,"",plaintext,content
+193,926010,"Untitled-1",44,4,"",plaintext,content
+194,928254,"Untitled-1",44,0,"    ",plaintext,content
+195,928611,"Untitled-1",48,0,"e",plaintext,content
+196,928611,"Untitled-1",49,0,"",plaintext,selection_keyboard
+197,928796,"Untitled-1",49,0,"l",plaintext,content
+198,928796,"Untitled-1",50,0,"",plaintext,selection_keyboard
+199,929911,"Untitled-1",44,7,"    else\n",plaintext,content
+200,930890,"Untitled-1",44,9,"    else:\n",plaintext,content
+201,931866,"Untitled-1",54,31,"        return fib(n-1) + fib(n-2)\n",plaintext,content
+202,936421,"Untitled-1",52,0,"s",plaintext,content
+203,936421,"Untitled-1",53,0,"",plaintext,selection_keyboard
+204,936765,"Untitled-1",53,0,"l",plaintext,content
+205,936765,"Untitled-1",54,0,"",plaintext,selection_keyboard
+206,936898,"Untitled-1",54,0,"s",plaintext,content
+207,936898,"Untitled-1",55,0,"",plaintext,selection_keyboard
+208,937002,"Untitled-1",55,0,"l",plaintext,content
+209,937002,"Untitled-1",56,0,"",plaintext,selection_keyboard
+210,937155,"Untitled-1",56,0,"s",plaintext,content
+211,937155,"Untitled-1",57,0,"",plaintext,selection_keyboard
+212,937156,"Untitled-1",57,0,"f",plaintext,content
+213,937156,"Untitled-1",58,0,"",plaintext,selection_keyboard
+214,939159,"Untitled-1",44,16,"    elseslslsf:\n",plaintext,content
+215,939851,"Untitled-1",44,16,"    elseslslsfds:\n",plaintext,content
+216,941130,"Untitled-1",57,0,"",plaintext,selection_command
+217,943121,"Untitled-1",56,0,"",plaintext,selection_command
+218,943166,"Untitled-1",55,0,"",plaintext,selection_command
+219,943268,"Untitled-1",54,0,"",plaintext,selection_command
+220,943434,"Untitled-1",53,0,"",plaintext,selection_command
+221,943958,"Untitled-1",53,8,"",plaintext,content
+222,945313,"Untitled-1",44,10,"    elses:\n",plaintext,content
+223,945545,"Untitled-1",52,0,"",plaintext,selection_command
+224,946467,"Untitled-1",53,0,"",plaintext,selection_command
+225,955082,"Untitled-1",53,0,"hello",plaintext,content
+226,955083,"Untitled-1",58,0,"",plaintext,selection_keyboard
+227,956487,"Untitled-1",44,16,"    else\n",plaintext,content
+228,957593,"Untitled-1",44,9,"    else:\n",plaintext,content
+229,968064,"Untitled-1",44,0,"",plaintext,selection_command
+230,968149,"Untitled-1",27,0,"",plaintext,selection_command
+231,968285,"Untitled-1",12,0,"",plaintext,selection_command
+232,968419,"Untitled-1",0,0,"",plaintext,selection_command
+233,968860,"Untitled-1",0,11,"def fib(n):",plaintext,selection_command
+234,969140,"Untitled-1",0,89,"def fib(n):\n    if n <= 1:\n        return n\n    else:\n        return fib(n-1) + fib(n-2)\n",plaintext,selection_command
+235,969658,"Untitled-1",0,89,"",plaintext,content
+236,980510,"Untitled-1",0,0,"d",plaintext,content
+237,980511,"Untitled-1",1,0,"",plaintext,selection_keyboard
+238,980619,"Untitled-1",1,0,"e",plaintext,content
+239,980619,"Untitled-1",2,0,"",plaintext,selection_keyboard
+240,980763,"Untitled-1",2,0,"f",plaintext,content
+241,980763,"Untitled-1",3,0,"",plaintext,selection_keyboard
+242,980808,"Untitled-1",3,0," ",plaintext,content
+243,980808,"Untitled-1",4,0,"",plaintext,selection_keyboard
+244,981033,"Untitled-1",4,0,"f",plaintext,content
+245,981033,"Untitled-1",5,0,"",plaintext,selection_keyboard
+246,981061,"Untitled-1",5,0,"i",plaintext,content
+247,981061,"Untitled-1",6,0,"",plaintext,selection_keyboard
+248,981147,"Untitled-1",6,0,"b",plaintext,content
+249,981147,"Untitled-1",7,0,"",plaintext,selection_keyboard
+250,981225,"Untitled-1",7,0,"o",plaintext,content
+251,981225,"Untitled-1",8,0,"",plaintext,selection_keyboard
+252,983198,"Untitled-1",8,0,"n",plaintext,content
+253,983199,"Untitled-1",9,0,"",plaintext,selection_keyboard
+254,983337,"Untitled-1",9,0,"n",plaintext,content
+255,983337,"Untitled-1",10,0,"",plaintext,selection_keyboard
+256,983801,"Untitled-1",9,1,"",plaintext,content
+257,983958,"Untitled-1",9,0,"a",plaintext,content
+258,983958,"Untitled-1",10,0,"",plaintext,selection_keyboard
+259,984048,"Untitled-1",10,0,"c",plaintext,content
+260,984048,"Untitled-1",11,0,"",plaintext,selection_keyboard
+261,984208,"Untitled-1",11,0,"c",plaintext,content
+262,984208,"Untitled-1",12,0,"",plaintext,selection_keyboard
+263,985797,"Untitled-1",0,12,"def fibonacci\n",plaintext,content
+264,987029,"Untitled-1",13,1,"",plaintext,content
+265,987769,"Untitled-1",13,0,"()",plaintext,content
+266,987769,"Untitled-1",14,0,"",plaintext,selection_keyboard
+267,989496,"Untitled-1",0,15,"def fibonacci(n):\n",plaintext,content
+268,991025,"Untitled-1",18,0,"",plaintext,selection_command
+269,992074,"Untitled-1",18,0,"    if n <= 1:\n",plaintext,content
+270,992944,"Untitled-1",33,0,"        return n\n",plaintext,content
+271,993687,"Untitled-1",50,0,"    else:\n",plaintext,content
+272,994629,"Untitled-1",60,0,"        return fibonacci(n-1) + fibonacci(n-2)\n",plaintext,content
+273,998093,"Untitled-1",60,0,"",plaintext,selection_command
+274,998465,"Untitled-1",50,0,"",plaintext,selection_command
+275,999433,"Untitled-1",58,0,"hello",plaintext,content
+276,999433,"Untitled-1",63,0,"",plaintext,selection_keyboard
+277,1000718,"Untitled-1",50,15,"    else:\n",plaintext,content
+278,1779802,"Untitled-1",0,107,"",plaintext,content
+279,1779807,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",0,0,"",shellscript,tab
+280,4203396,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",21,0,"",shellscript,selection_command
+281,4208550,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",39,0,"",shellscript,selection_command
+282,4208692,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",67,0,"",shellscript,selection_command
+283,4224569,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",0,0,"",shellscript,tab
+284,4233455,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",21,0,"",shellscript,selection_command
+285,4234389,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",20,0,"",shellscript,selection_command
+286,4246610,"TERMINAL",0,0,"claude",,terminal_command
+287,4246731,"TERMINAL",0,0,"]633;C",,terminal_output
+288,4247723,"TERMINAL",0,0,"[?2026h[?25l\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n Do you trust the files in this folder?\r\n\r\n /fast/home/franz.srambical/crowd-pilot-extension\r\n\r\n Claude Code may read, write, or execute files contained in this directory. This can pose security risks, so only use files from trusted sources.\r\n\r\n ]8;;https://code.claude.com/docs/en/securityLearn more]8;;\r\n\r\n ❯ 1. Yes, proceed\r\n   2. No, exit\r\n\r\n Enter to confirm · Esc to cancel\r\n[?2026l[?25l",,terminal_output
+289,4247768,"TERMINAL",0,0,"[?2004h",,terminal_output
+290,4250161,"TERMINAL",0,0,"[?2026h[?25h[?2026l[?25h[?2004l",,terminal_output
+291,4250409,"TERMINAL",0,0,"[?25l[?2004h",,terminal_output
+292,4250656,"TERMINAL",0,0,"[?25h[?2004l",,terminal_output
+293,4250722,"TERMINAL",0,0,"[?2026h\r\n╭─── Claude Code v2.0.76 ────��───────────────────────────────────────────────────────────────────────────────────────────╮\r\n│                                                    │ Tips for getting started                                          │\r\n│                Welcome back Kalyan!                │ Run /init to create a CLAUDE.md file with instructions for Claude │\r\n│                                                    │ ───────────────────────────────────────────────────────────────── │\r\n│                     * ▐▛███▜▌ *                    │ Recent activity                                                   │\r\n│                    * ▝▜█████▛▘ *                   │ No recent activity                                                │\r\n│                     *  ▘▘ ▝▝  *                    │                                                                   │\r\n│                                                    │                                                                   │\r\n│   Opus 4.5 · Claude Max · nkalyanv@gmail.com's     │                                                                   │\r\n│   Organization                                     │                                                                   │\r\n│  /fast/home/franz.srambical/crowd-pilot-extension  │                                                                   │\r\n╰────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\r\n\r\n\r\n You now have access to Sonnet 4.5 with 1M context (uses more rate limits than Sonnet on long requests) • Update in /model\r\n[?25l\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n> Try ""create a util logging.py that...""\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l[?25l",,terminal_output
+294,4250775,"TERMINAL",0,0,"[?2004h[?1004h",,terminal_output
+295,4252227,"TERMINAL",0,0,"[?2026h\r\n───────��──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n> Try ""create a util logging.py that...""\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output
+296,4254520,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",0,0,"",shellscript,selection_command
+297,4256443,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",20,0,"",shellscript,selection_command
+298,4258412,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n> Try ""create a util logging.py that...""\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output
+299,4266768,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n> / \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n[?2026l[?2026h\r\n───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────���──────────────────────────────────────────\r\n> / \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  /add-dir                Add a new working directory\r\n  /agents                 Manage agent configurations\r\n  /chrome                 Claude in Chrome (Beta) settings\r\n  /clear                  Clear conversation history and free up context\r\n  /compact                Clear conversation history but keep a summary in context. Optional: /compact [instructions for summarization]\r\n  /config                 Open config panel\r\n  /context                Visualize current context usage as a colored grid\r\n  /doctor                 Diagnose and verify your Claude Code installation and settings\r\n  /exit                   Exit the REPL\r\n  /export                 Export the current conversation to a file or clipboard\r\n[?2026l",,terminal_output
+300,4266877,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n> /m \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  /mcp                      Manage MCP servers\r\n  /memory                   Edit Claude memory files\r\n  /mobile                   Show QR code to download the Claude mobile app\r\n  /model                    Set the AI model for Claude Code\r\n  /remote-env               Configure the default remote environment for teleport sessions\r\n  /pr-comments              Get comments from a GitHub pull request\r\n  /vim                      Toggle between Vim and Normal editing modes\r\n  /compact                  Clear conversation history but keep a summary in context. Optional: /compact [instructions for summarization]\r\n  /terminal-setup           Install Shift+Enter key binding for newlines\r\n  /plugin (marketplace)     Manage Claude Code plugins\r\n[?2026l",,terminal_output
+301,4266939,"TERMINAL",0,0,"[?2026h\r\n─────────────────────────────────────────────────────���────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n> /mo \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  /mobile         Show QR code to download the Claude mobile app\r\n  /model          Set the AI model for Claude Code\r\n  /remote-env     Configure the default remote environment for teleport sessions\r\n  /memory         Edit Claude memory files\r\n  /status         Show Claude Code status including version, model, account, API connectivity, and tool statuses\r\n  /upgrade        Upgrade to Max for higher rate limits and more Opus\r\n  /vim            Toggle between Vim and Normal editing modes\r\n[?2026l",,terminal_output
+302,4267073,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n> /mod \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  /model      Set the AI model for Claude Code\r\n  /status     Show Claude Code status including version, model, account, API connectivity, and tool statuses\r\n  /vim        Toggle between Vim and Normal editing modes\r\n[?2026l",,terminal_output
+303,4267213,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n> /mode \r\n───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────��──────────────────────────────────\r\n  /model          Set the AI model for Claude Code\r\n  /remote-env     Configure the default remote environment for teleport sessions\r\n  /status         Show Claude Code status including version, model, account, API connectivity, and tool statuses\r\n  /vim            Toggle between Vim and Normal editing modes\r\n  /passes         Share a free week of Claude Code with friends\r\n  /doctor         Diagnose and verify your Claude Code installation and settings\r\n  /init           Initialize a new CLAUDE.md file with codebase documentation\r\n  /mobile         Show QR code to download the Claude mobile app\r\n  /plugin         Manage Claude Code plugins\r\n  /stats          Show your Claude Code usage statistics and activity\r\n[?2026l",,terminal_output
+304,4267386,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n> /model  [model]\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  /model      Set the AI model for Claude Code\r\n  /status     Show Claude Code status including version, model, account, API connectivity, and tool statuses\r\n  /vim        Toggle between Vim and Normal editing modes\r\n[?2026l",,terminal_output
+305,4272455,"TERMINAL",0,0,"[?2026h\r\n> /model \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n Select model\r\n Switch between Claude models. Applies to this session and future Claude Code sessions. For other/previous model names, specify with --model.\r\n\r\n ❯ 1. Default (recommended) ✔  Opus 4.5 · Most capable for complex work\r\n   2. Sonnet                   Sonnet 4.5 · Best for everyday tasks\r\n   3. Sonnet (1M context)      Sonnet 4.5 with 1M context · Uses rate limits faster\r\n   4. Haiku                    Haiku 4.5 · Fastest for quick answers\r\n\r\n Enter to confirm · Esc to exit\r\n[?2026l[?1004l",,terminal_output
+306,4278893,"TERMINAL",0,0,"[?2026h\r\n> /model \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n Select model\r\n Switch between Claude models. Applies to this session and future Claude Code sessions. For other/previous model names, specify with --model.\r\n\r\n   1. Default (recommended) ✔  Opus 4.5 · Most capable for complex work\r\n ❯ 2. Sonnet                   Sonnet 4.5 · Best for everyday tasks\r\n   3. Sonnet (1M context)      Sonnet 4.5 with 1M context · Uses rate limits faster\r\n   4. Haiku                    Haiku 4.5 · Fastest for quick answers\r\n\r\n Enter to confirm · Esc to exit\r\n[?2026l",,terminal_output
+307,4279006,"TERMINAL",0,0,"[?2026h\r\n> /model \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n Select model\r\n Switch between Claude models. Applies to this session and future Claude Code sessions. For other/previous model names, specify with --model.\r\n\r\n   1. Default (recommended) ✔  Opus 4.5 · Most capable for complex work\r\n   2. Sonnet                   Sonnet 4.5 · Best for everyday tasks\r\n ❯ 3. Sonnet (1M context)      Sonnet 4.5 with 1M context · Uses rate limits faster\r\n   4. Haiku                    Haiku 4.5 · Fastest for quick answers\r\n\r\n Enter to confirm · Esc to exit\r\n[?2026l",,terminal_output
+308,4279227,"TERMINAL",0,0,"[?2026h\r\n> /model \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n Select model\r\n Switch between Claude models. Applies to this session and future Claude Code sessions. For other/previous model names, specify with --model.\r\n\r\n   1. Default (recommended) ✔  Opus 4.5 · Most capable for complex work\r\n ❯ 2. Sonnet                   Sonnet 4.5 · Best for everyday tasks\r\n   3. Sonnet (1M context)      Sonnet 4.5 with 1M context · Uses rate limits faster\r\n   4. Haiku                    Haiku 4.5 · Fastest for quick answers\r\n\r\n Enter to confirm · Esc to exit\r\n[?2026l",,terminal_output
+309,4283702,"TERMINAL",0,0,"[?2026h\r\n> /model \r\n  ⎿  Set model to sonnet (claude-sonnet-4-5-20250929)\r\n\r\n· Ruminating… (esc to interrupt)\r\n\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>  \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l[?1004h[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>  \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output
+310,4283758,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>  \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts                                                                                                                                                                   Checking for updates\r\n[?2026l",,terminal_output
+311,4284049,"TERMINAL",0,0,"[?2026h\r\n─────────────────────────────────────────────────��────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>  \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output
+312,4287599,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output
+313,4342294,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>  \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output
+314,4344322,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>\r\n─────────────────────────────────────────────────────────────────────────────────���────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output
+315,4344871,"TERMINAL",0,0,"",,terminal_focus
+316,4345025,"TERMINAL",0,0,"[?2026h\r\n╭─── Claude Code v2.0.76 ────────────────────────────────────────────────────────────────────────────────────────────────╮\r\n│                                                    │ Tips for getting started                                          │\r\n│                Welcome back Kalyan!                │ Run /init to create a CLAUDE.md file with instructions for Claude │\r\n│                                                    │ ───────────────────────────────────────────────────────────────── │\r\n│                     * ▐▛███▜▌ *                    │ Recent activity                                                   │\r\n│                    * ▝▜█████▛▘ *                   │ No recent activity                                                │\r\n│                     *  ▘▘ ▝▝  *                    │                                                                   │\r\n│                                                    │                                                                   │\r\n│   Opus 4.5 · Claude Max · nkalyanv@gmail.com's     │                                                                   │\r\n│   Organization                                     │                                                                   │\r\n│  /fast/home/franz.srambical/crowd-pilot-extension  │                                                                   │\r\n╰────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯\r\n\r\n\r\n You now have access to Sonnet 4.5 with 1M context (uses more rate limits than Sonnet on long requests) • Update in /model\r\n\r\n> /model \r\n  ⎿  Set model to sonnet (claude-sonnet-4-5-20250929)\r\n\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>  \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output
+317,4350563,"src/preview/quickPick.ts",0,0,"",typescript,tab
+318,4353229,"/home/franz.srambical/slurm/jobs/franz/berlin/crowd-pilot/start_sglang_server_glm4_5_air.sh",0,0,"",shellscript,tab
+319,4356154,"package.json",0,0,"",json,tab
+320,4358030,"package.json",1365,0,"",json,selection_keyboard
+321,4358420,"package.json",3275,0,"",json,selection_keyboard
+322,4358887,"package.json",4934,0,"",json,selection_keyboard
+323,4359844,"package.json",4036,0,"",json,selection_keyboard
+324,4374998,"package.json",4008,0,"",json,selection_command
+325,4375152,"package.json",3964,0,"",json,selection_command
+326,4375310,"package.json",3949,0,"",json,selection_command
+327,4375547,"package.json",3943,0,"",json,selection_command
+328,4375571,"package.json",3938,0,"",json,selection_command
+329,4375613,"package.json",3930,0,"",json,selection_command
+330,4375638,"package.json",3867,0,"",json,selection_command
+331,4375765,"package.json",3833,0,"",json,selection_command
+332,4375991,"package.json",3798,0,"",json,selection_command
+333,4380093,"package.json",3794,34,"        ""key"": ""ctrl+shift+space"",",json,selection_command
+334,4380285,"package.json",3794,68,"        ""key"": ""ctrl+shift+space"",\n        ""mac"": ""cmd+shift+space"",",json,selection_command
+335,4382055,"package.json",3833,0,"",json,selection_command
+336,4385132,"TERMINAL",0,0,"claude",,terminal_focus
+337,4385824,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>  \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output
+338,4388052,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output
+339,4464591,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>  \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output
+340,4465474,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output
+341,4494735,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>  \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output
+342,4495134,"TERMINAL",0,0,"[?2026h\r\n─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────��────────────────────────────────────────────────────────────────\r\n>\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output
+343,4497108,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>  \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output
+344,4497933,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output
+345,4499553,"TERMINAL",0,0,"[?2026h\r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n>  \r\n──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────\r\n  ? for shortcuts\r\n[?2026l",,terminal_output

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-72e8191a-80be-4131-b710-cd43dc0f498d1758216275170-2025_09_18-19.24.57.192/source.csv

@@ -0,0 +1,9 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"cleanrl/ppo_atari_envpool.py",0,0,"# docs and experiment results can be found at https://docs.cleanrl.dev/rl-algorithms/ppo/#ppo_atari_envpoolpy\nimport os\nimport random\nimport time\nfrom collections import deque\nfrom dataclasses import dataclass\n\nimport envpool\nimport gym\nimport numpy as np\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\nimport tyro\nfrom torch.distributions.categorical import Categorical\nfrom torch.utils.tensorboard import SummaryWriter\n\n\n@dataclass\nclass Args:\n    exp_name: str = os.path.basename(__file__)[: -len("".py"")]\n    """"""the name of this experiment""""""\n    seed: int = 1\n    """"""seed of the experiment""""""\n    torch_deterministic: bool = True\n    """"""if toggled, `torch.backends.cudnn.deterministic=False`""""""\n    cuda: bool = True\n    """"""if toggled, cuda will be enabled by default""""""\n    track: bool = False\n    """"""if toggled, this experiment will be tracked with Weights and Biases""""""\n    wandb_project_name: str = ""cleanRL""\n    """"""the wandb's project name""""""\n    wandb_entity: str = None\n    """"""the entity (team) of wandb's project""""""\n    capture_video: bool = False\n    """"""whether to capture videos of the agent performances (check out `videos` folder)""""""\n\n    # Algorithm specific arguments\n    env_id: str = ""Breakout-v5""\n    """"""the id of the environment""""""\n    total_timesteps: int = 10000000\n    """"""total timesteps of the experiments""""""\n    learning_rate: float = 2.5e-4\n    """"""the learning rate of the optimizer""""""\n    num_envs: int = 8\n    """"""the number of parallel game environments""""""\n    num_steps: int = 128\n    """"""the number of steps to run in each environment per policy rollout""""""\n    anneal_lr: bool = True\n    """"""Toggle learning rate annealing for policy and value networks""""""\n    gamma: float = 0.99\n    """"""the discount factor gamma""""""\n    gae_lambda: float = 0.95\n    """"""the lambda for the general advantage estimation""""""\n    num_minibatches: int = 4\n    """"""the number of mini-batches""""""\n    update_epochs: int = 4\n    """"""the K epochs to update the policy""""""\n    norm_adv: bool = True\n    """"""Toggles advantages normalization""""""\n    clip_coef: float = 0.1\n    """"""the surrogate clipping coefficient""""""\n    clip_vloss: bool = True\n    """"""Toggles whether or not to use a clipped loss for the value function, as per the paper.""""""\n    ent_coef: float = 0.01\n    """"""coefficient of the entropy""""""\n    vf_coef: float = 0.5\n    """"""coefficient of the value function""""""\n    max_grad_norm: float = 0.5\n    """"""the maximum norm for the gradient clipping""""""\n    target_kl: float = None\n    """"""the target KL divergence threshold""""""\n\n    # to be filled in runtime\n    batch_size: int = 0\n    """"""the batch size (computed in runtime)""""""\n    minibatch_size: int = 0\n    """"""the mini-batch size (computed in runtime)""""""\n    num_iterations: int = 0\n    """"""the number of iterations (computed in runtime)""""""\n\n\nclass RecordEpisodeStatistics(gym.Wrapper):\n    def __init__(self, env, deque_size=100):\n        super().__init__(env)\n        self.num_envs = getattr(env, ""num_envs"", 1)\n        self.episode_returns = None\n        self.episode_lengths = None\n\n    def reset(self, **kwargs):\n        observations = super().reset(**kwargs)\n        self.episode_returns = np.zeros(self.num_envs, dtype=np.float32)\n        self.episode_lengths = np.zeros(self.num_envs, dtype=np.int32)\n        self.lives = np.zeros(self.num_envs, dtype=np.int32)\n        self.returned_episode_returns = np.zeros(self.num_envs, dtype=np.float32)\n        self.returned_episode_lengths = np.zeros(self.num_envs, dtype=np.int32)\n        return observations\n\n    def step(self, action):\n        observations, rewards, dones, infos = super().step(action)\n        self.episode_returns += infos[""reward""]\n        self.episode_lengths += 1\n        self.returned_episode_returns[:] = self.episode_returns\n        self.returned_episode_lengths[:] = self.episode_lengths\n        self.episode_returns *= 1 - infos[""terminated""]\n        self.episode_lengths *= 1 - infos[""terminated""]\n        infos[""r""] = self.returned_episode_returns\n        infos[""l""] = self.returned_episode_lengths\n        return (\n            observations,\n            rewards,\n            dones,\n            infos,\n        )\n\n\ndef layer_init(layer, std=np.sqrt(2), bias_const=0.0):\n    torch.nn.init.orthogonal_(layer.weight, std)\n    torch.nn.init.constant_(layer.bias, bias_const)\n    return layer\n\n\nclass Agent(nn.Module):\n    def __init__(self, envs):\n        super().__init__()\n        self.network = nn.Sequential(\n            layer_init(nn.Conv2d(4, 32, 8, stride=4)),\n            nn.ReLU(),\n            layer_init(nn.Conv2d(32, 64, 4, stride=2)),\n            nn.ReLU(),\n            layer_init(nn.Conv2d(64, 64, 3, stride=1)),\n            nn.ReLU(),\n            nn.Flatten(),\n            layer_init(nn.Linear(64 * 7 * 7, 512)),\n            nn.ReLU(),\n        )\n        self.actor = layer_init(nn.Linear(512, envs.single_action_space.n), std=0.01)\n        self.critic = layer_init(nn.Linear(512, 1), std=1)\n\n    def get_value(self, x):\n        return self.critic(self.network(x / 255.0))\n\n    def get_action_and_value(self, x, action=None):\n        hidden = self.network(x / 255.0)\n        logits = self.actor(hidden)\n        probs = Categorical(logits=logits)\n        if action is None:\n            action = probs.sample()\n        return action, probs.log_prob(action), probs.entropy(), self.critic(hidden)\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(Args)\n    args.batch_size = int(args.num_envs * args.num_steps)\n    args.minibatch_size = int(args.batch_size // args.num_minibatches)\n    args.num_iterations = args.total_timesteps // args.batch_size\n    run_name = f""{args.env_id}__{args.exp_name}__{args.seed}__{int(time.time())}""\n    if args.track:\n        import wandb\n\n        wandb.init(\n            project=args.wandb_project_name,\n            entity=args.wandb_entity,\n            sync_tensorboard=True,\n            config=vars(args),\n            name=run_name,\n            monitor_gym=True,\n            save_code=True,\n        )\n    writer = SummaryWriter(f""runs/{run_name}"")\n    writer.add_text(\n        ""hyperparameters"",\n        ""|param|value|\n|-|-|\n%s"" % (""\n"".join([f""|{key}|{value}|"" for key, value in vars(args).items()])),\n    )\n\n    # TRY NOT TO MODIFY: seeding\n    random.seed(args.seed)\n    np.random.seed(args.seed)\n    torch.manual_seed(args.seed)\n    torch.backends.cudnn.deterministic = args.torch_deterministic\n\n    device = torch.device(""cuda"" if torch.cuda.is_available() and args.cuda else ""cpu"")\n\n    # env setup\n    envs = envpool.make(\n        args.env_id,\n        env_type=""gym"",\n        num_envs=args.num_envs,\n        episodic_life=True,\n        reward_clip=True,\n        seed=args.seed,\n    )\n    envs.num_envs = args.num_envs\n    envs.single_action_space = envs.action_space\n    envs.single_observation_space = envs.observation_space\n    envs = RecordEpisodeStatistics(envs)\n    assert isinstance(envs.action_space, gym.spaces.Discrete), ""only discrete action space is supported""\n\n    agent = Agent(envs).to(device)\n    optimizer = optim.Adam(agent.parameters(), lr=args.learning_rate, eps=1e-5)\n\n    # ALGO Logic: Storage setup\n    obs = torch.zeros((args.num_steps, args.num_envs) + envs.single_observation_space.shape).to(device)\n    actions = torch.zeros((args.num_steps, args.num_envs) + envs.single_action_space.shape).to(device)\n    logprobs = torch.zeros((args.num_steps, args.num_envs)).to(device)\n    rewards = torch.zeros((args.num_steps, args.num_envs)).to(device)\n    dones = torch.zeros((args.num_steps, args.num_envs)).to(device)\n    values = torch.zeros((args.num_steps, args.num_envs)).to(device)\n    avg_returns = deque(maxlen=20)\n\n    # TRY NOT TO MODIFY: start the game\n    global_step = 0\n    start_time = time.time()\n    next_obs = torch.Tensor(envs.reset()).to(device)\n    next_done = torch.zeros(args.num_envs).to(device)\n\n    for iteration in range(1, args.num_iterations + 1):\n        # Annealing the rate if instructed to do so.\n        if args.anneal_lr:\n            frac = 1.0 - (iteration - 1.0) / args.num_iterations\n            lrnow = frac * args.learning_rate\n            optimizer.param_groups[0][""lr""] = lrnow\n\n        for step in range(0, args.num_steps):\n            global_step += args.num_envs\n            obs[step] = next_obs\n            dones[step] = next_done\n\n            # ALGO LOGIC: action logic\n            with torch.no_grad():\n                action, logprob, _, value = agent.get_action_and_value(next_obs)\n                values[step] = value.flatten()\n            actions[step] = action\n            logprobs[step] = logprob\n\n            # TRY NOT TO MODIFY: execute the game and log data.\n            next_obs, reward, next_done, info = envs.step(action.cpu().numpy())\n            rewards[step] = torch.tensor(reward).to(device).view(-1)\n            next_obs, next_done = torch.Tensor(next_obs).to(device), torch.Tensor(next_done).to(device)\n\n            for idx, d in enumerate(next_done):\n                if d and info[""lives""][idx] == 0:\n                    print(f""global_step={global_step}, episodic_return={info['r'][idx]}"")\n                    avg_returns.append(info[""r""][idx])\n                    writer.add_scalar(""charts/avg_episodic_return"", np.average(avg_returns), global_step)\n                    writer.add_scalar(""charts/episodic_return"", info[""r""][idx], global_step)\n                    writer.add_scalar(""charts/episodic_length"", info[""l""][idx], global_step)\n\n        # bootstrap value if not done\n        with torch.no_grad():\n            next_value = agent.get_value(next_obs).reshape(1, -1)\n            advantages = torch.zeros_like(rewards).to(device)\n            lastgaelam = 0\n            for t in reversed(range(args.num_steps)):\n                if t == args.num_steps - 1:\n                    nextnonterminal = 1.0 - next_done\n                    nextvalues = next_value\n                else:\n                    nextnonterminal = 1.0 - dones[t + 1]\n                    nextvalues = values[t + 1]\n                delta = rewards[t] + args.gamma * nextvalues * nextnonterminal - values[t]\n                advantages[t] = lastgaelam = delta + args.gamma * args.gae_lambda * nextnonterminal * lastgaelam\n            returns = advantages + values\n\n        # flatten the batch\n        b_obs = obs.reshape((-1,) + envs.single_observation_space.shape)\n        b_logprobs = logprobs.reshape(-1)\n        b_actions = actions.reshape((-1,) + envs.single_action_space.shape)\n        b_advantages = advantages.reshape(-1)\n        b_returns = returns.reshape(-1)\n        b_values = values.reshape(-1)\n\n        # Optimizing the policy and value network\n        b_inds = np.arange(args.batch_size)\n        clipfracs = []\n        for epoch in range(args.update_epochs):\n            np.random.shuffle(b_inds)\n            for start in range(0, args.batch_size, args.minibatch_size):\n                end = start + args.minibatch_size\n                mb_inds = b_inds[start:end]\n\n                _, newlogprob, entropy, newvalue = agent.get_action_and_value(b_obs[mb_inds], b_actions.long()[mb_inds])\n                logratio = newlogprob - b_logprobs[mb_inds]\n                ratio = logratio.exp()\n\n                with torch.no_grad():\n                    # calculate approx_kl http://joschu.net/blog/kl-approx.html\n                    old_approx_kl = (-logratio).mean()\n                    approx_kl = ((ratio - 1) - logratio).mean()\n                    clipfracs += [((ratio - 1.0).abs() > args.clip_coef).float().mean().item()]\n\n                mb_advantages = b_advantages[mb_inds]\n                if args.norm_adv:\n                    mb_advantages = (mb_advantages - mb_advantages.mean()) / (mb_advantages.std() + 1e-8)\n\n                # Policy loss\n                pg_loss1 = -mb_advantages * ratio\n                pg_loss2 = -mb_advantages * torch.clamp(ratio, 1 - args.clip_coef, 1 + args.clip_coef)\n                pg_loss = torch.max(pg_loss1, pg_loss2).mean()\n\n                # Value loss\n                newvalue = newvalue.view(-1)\n                if args.clip_vloss:\n                    v_loss_unclipped = (newvalue - b_returns[mb_inds]) ** 2\n                    v_clipped = b_values[mb_inds] + torch.clamp(\n                        newvalue - b_values[mb_inds],\n                        -args.clip_coef,\n                        args.clip_coef,\n                    )\n                    v_loss_clipped = (v_clipped - b_returns[mb_inds]) ** 2\n                    v_loss_max = torch.max(v_loss_unclipped, v_loss_clipped)\n                    v_loss = 0.5 * v_loss_max.mean()\n                else:\n                    v_loss = 0.5 * ((newvalue - b_returns[mb_inds]) ** 2).mean()\n\n                entropy_loss = entropy.mean()\n                loss = pg_loss - args.ent_coef * entropy_loss + v_loss * args.vf_coef\n\n                optimizer.zero_grad()\n                loss.backward()\n                nn.utils.clip_grad_norm_(agent.parameters(), args.max_grad_norm)\n                optimizer.step()\n\n            if args.target_kl is not None and approx_kl > args.target_kl:\n                break\n\n        y_pred, y_true = b_values.cpu().numpy(), b_returns.cpu().numpy()\n        var_y = np.var(y_true)\n        explained_var = np.nan if var_y == 0 else 1 - np.var(y_true - y_pred) / var_y\n\n        # TRY NOT TO MODIFY: record rewards for plotting purposes\n        writer.add_scalar(""charts/learning_rate"", optimizer.param_groups[0][""lr""], global_step)\n        writer.add_scalar(""losses/value_loss"", v_loss.item(), global_step)\n        writer.add_scalar(""losses/policy_loss"", pg_loss.item(), global_step)\n        writer.add_scalar(""losses/entropy"", entropy_loss.item(), global_step)\n        writer.add_scalar(""losses/old_approx_kl"", old_approx_kl.item(), global_step)\n        writer.add_scalar(""losses/approx_kl"", approx_kl.item(), global_step)\n        writer.add_scalar(""losses/clipfrac"", np.mean(clipfracs), global_step)\n        writer.add_scalar(""losses/explained_variance"", explained_var, global_step)\n        print(""SPS:"", int(global_step / (time.time() - start_time)))\n        writer.add_scalar(""charts/SPS"", int(global_step / (time.time() - start_time)), global_step)\n\n    envs.close()\n    writer.close()\n",python,tab
+2,1469,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"7:24:57 PM [info] Activating crowd-code\n7:24:57 PM [info] Recording started\n7:24:57 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,1633,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"7:24:58 PM [info] Git repository found\n7:24:58 PM [info] Git provider initialized successfully\n7:24:58 PM [info] Initial git state: [object Object]\n",Log,content
+4,1636,"extension-output-pdoom-org.crowd-code-#1-crowd-code",298,0,"",Log,selection_mouse
+5,3292,"cleanrl/ppo_atari_envpool.py",0,0,"",python,tab
+6,4347,"TERMINAL",0,0,"squeue",,terminal_command
+7,4369,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  29260   xiao.liu interacti     1    64  R 2025-09-18T14:49:19 2025-09-18T14:49:19  4:35:42   23:59:00 hai004\r\n  29258   xiao.liu interacti     1    64  R 2025-09-18T11:19:00 2025-09-18T11:19:00  8:06:01   23:59:00 hai005\r\n]0;franz.srambical@hai-login2:~/cleanrl",,terminal_output
+8,13313,"TERMINAL",0,0,"",,terminal_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-737b410d-69c4-4399-9585-9581396eacb11765644421874-2025_12_13-17.48.18.686/source.csv

@@ -0,0 +1,199 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"crates/core/src/conversation.rs",0,0,"//! Conversation state manager for serializing IDE events into conversation format.\n\nuse std::collections::{HashMap, HashSet};\n\nuse crate::diff::compute_changed_block_lines;\nuse crate::helpers::{\n    clean_text, escape_single_quotes_for_sed, fenced_block, line_numbered_output,\n    normalize_terminal_output, serialize_compute_viewport, Viewport,\n};\nuse crate::Tokenizer;\nuse crate::{COALESCE_RADIUS, MAX_TOKENS_PER_MESSAGE, MAX_TOKENS_PER_TERMINAL_OUTPUT, VIEWPORT_RADIUS};\n\n/// A single message in the conversation.\n#[derive(Debug, Clone, PartialEq, Eq)]\npub struct ConversationMessage {\n    pub from: String,\n    pub value: String,\n}\n\nimpl ConversationMessage {\n    pub fn user(value: impl Into<String>) -> Self {\n        Self {\n            from: ""User"".to_string(),\n            value: value.into(),\n        }\n    }\n\n    pub fn assistant(value: impl Into<String>) -> Self {\n        Self {\n            from: ""Assistant"".to_string(),\n            value: value.into(),\n        }\n    }\n}\n\n/// Configuration for the ConversationStateManager.\n#[derive(Debug, Clone)]\npub struct ConversationStateManagerConfig {\n    pub viewport_radius: usize,\n    pub coalesce_radius: usize,\n    pub max_tokens_per_message: usize,\n    pub max_tokens_per_terminal_output: usize,\n    /// Maximum tokens per conversation chunk (for preprocessing). None = no chunking.\n    pub max_tokens_per_conversation: Option<usize>,\n    /// Minimum messages required to keep a conversation chunk.\n    pub min_conversation_messages: usize,\n}\n\nimpl Default for ConversationStateManagerConfig {\n    fn default() -> Self {\n        Self {\n            viewport_radius: VIEWPORT_RADIUS,\n            coalesce_radius: COALESCE_RADIUS,\n            max_tokens_per_message: MAX_TOKENS_PER_MESSAGE,\n            max_tokens_per_terminal_output: MAX_TOKENS_PER_TERMINAL_OUTPUT,\n            max_tokens_per_conversation: None, // No chunking by default (for extension)\n            min_conversation_messages: 5,\n        }\n    }\n}\n\n/// A finalized conversation with its token count.\n#[derive(Debug, Clone)]\npub struct FinalizedConversation {\n    pub messages: Vec<ConversationMessage>,\n    pub token_count: usize,\n}\n\n/// Edit region tracking for coalescing nearby edits.\n#[derive(Debug, Clone, Copy)]\nstruct EditRegion {\n    start: usize,\n    end: usize,\n}\n\n/// Manages conversation state for serializing IDE events.\n///\n/// The tokenizer is provided externally, allowing the caller to use\n/// either a character-based approximation (for runtime) or an accurate tokenizer\n/// (for preprocessing).\npub struct ConversationStateManager<T>\nwhere\n    T: Tokenizer,\n{\n    tokenizer: T,\n    config: ConversationStateManagerConfig,\n    // Current conversation being built\n    messages: Vec<ConversationMessage>,\n    current_tokens: usize,\n    // Finalized conversations (for chunking mode)\n    finalized_conversations: Vec<FinalizedConversation>,\n    // File state tracking\n    file_states: HashMap<String, String>,\n    per_file_viewport: HashMap<String, Option<Viewport>>,\n    files_opened_in_conversation: HashSet<String>,\n    terminal_output_buffer: Vec<String>,\n    pending_edits_before: HashMap<String, Option<String>>,\n    pending_edit_regions: HashMap<String, Option<EditRegion>>,\n}\n\nimpl<T> ConversationStateManager<T>\nwhere\n    T: Tokenizer,\n{\n    /// Create a new ConversationStateManager with the given tokenizer.\n    pub fn new(tokenizer: T, config: ConversationStateManagerConfig) -> Self {\n        Self {\n            tokenizer,\n            config,\n            messages: Vec::new(),\n            current_tokens: 0,\n            finalized_conversations: Vec::new(),\n            file_states: HashMap::new(),\n            per_file_viewport: HashMap::new(),\n            files_opened_in_conversation: HashSet::new(),\n            terminal_output_buffer: Vec::new(),\n            pending_edits_before: HashMap::new(),\n            pending_edit_regions: HashMap::new(),\n        }\n    }\n\n    /// Reset all state.\n    pub fn reset(&mut self) {\n        self.messages.clear();\n        self.current_tokens = 0;\n        self.finalized_conversations.clear();\n        self.file_states.clear();\n        self.per_file_viewport.clear();\n        self.files_opened_in_conversation.clear();\n        self.terminal_output_buffer.clear();\n        self.pending_edits_before.clear();\n        self.pending_edit_regions.clear();\n    }\n\n    /// Finalize the current conversation and start a new one.\n    /// This is called automatically when conversation token limit is exceeded.\n    fn finalize_current_conversation(&mut self) {\n        if self.messages.is_empty() {\n            return;\n        }\n\n        // Check if conversation meets minimum requirements\n        let is_long_enough = self.messages.len() >= self.config.min_conversation_messages;\n        let has_user = self.messages.iter().any(|m| m.from == ""User"");\n        let has_assistant = self.messages.iter().any(|m| m.from == ""Assistant"");\n\n        if is_long_enough && has_user && has_assistant {\n            self.finalized_conversations.push(FinalizedConversation {\n                messages: std::mem::take(&mut self.messages),\n                token_count: self.current_tokens,\n            });\n        } else {\n            self.messages.clear();\n        }\n\n        self.current_tokens = 0;\n        self.files_opened_in_conversation.clear();\n    }\n\n    /// Get all finalized conversations with their token counts.\n    /// Call this after processing all events.\n    pub fn get_conversations(&mut self) -> Vec<FinalizedConversation> {\n        // Finalize any remaining conversation\n        self.flush_all_pending_edits();\n        self.flush_terminal_output_buffer();\n        self.finalize_current_conversation();\n        \n        std::mem::take(&mut self.finalized_conversations)\n    }\n\n    /// Get a copy of all messages.\n    pub fn get_messages(&self) -> Vec<ConversationMessage> {\n        self.messages.clone()\n    }\n\n    /// Get the current content of a file.\n    pub fn get_file_content(&self, file_path: &str) -> String {\n        self.file_states.get(file_path).cloned().unwrap_or_default()\n    }\n\n    /// Append a message, truncating if it exceeds token limits.\n    /// If chunking is enabled and conversation limit would be exceeded,\n    /// finalizes current conversation and starts a new one.\n    fn append_message(&mut self, mut message: ConversationMessage) {\n        let mut tokens = self.tokenizer.count_tokens(&message.value);\n        \n        if tokens > self.config.max_tokens_per_message {\n            message.value = self.tokenizer.truncate_to_max_tokens(\n                &message.value,\n                self.config.max_tokens_per_message,\n            );\n            tokens = self.config.max_tokens_per_message;\n        }\n\n        // Check if we need to start a new conversation (chunking mode)\n        if let Some(max_tokens) = self.config.max_tokens_per_conversation {\n            if self.current_tokens + tokens > max_tokens && !self.messages.is_empty() {\n                self.finalize_current_conversation();\n                // After starting a new conversation, we need to re-capture file states\n                // This will happen naturally as files are accessed\n            }\n        }\n\n        self.messages.push(message);\n        self.current_tokens += tokens;\n    }\n\n    /// Capture file contents if not already shown in this conversation.\n    fn maybe_capture_file_contents(&mut self, file_path: &str, content: &str) {\n        if self.files_opened_in_conversation.contains(file_path) {\n            return;\n        }\n        let cmd = format!(""cat -n {}"", file_path);\n        self.append_message(ConversationMessage::assistant(fenced_block(\n            Some(""bash""),\n            &clean_text(&cmd),\n        )));\n        let output = line_numbered_output(content, None, None);\n        self.append_message(ConversationMessage::user(format!(\n            ""<stdout>\n{}\n</stdout>"",\n            output\n        )));\n        self.files_opened_in_conversation.insert(file_path.to_string());\n    }\n\n    /// Flush buffered terminal output.\n    pub fn flush_terminal_output_buffer(&mut self) {\n        if self.terminal_output_buffer.is_empty() {\n            return;\n        }\n        let aggregated: String = self.terminal_output_buffer.join("""");\n        let out = normalize_terminal_output(&aggregated);\n        let mut cleaned = clean_text(&out);\n\n        let tokens = self.tokenizer.count_tokens(&cleaned);\n        if tokens > self.config.max_tokens_per_terminal_output {\n            let truncated = self.tokenizer.truncate_to_max_tokens(\n                &cleaned,\n                self.config.max_tokens_per_terminal_output,\n            );\n            cleaned = format!(""{}\n... [truncated]"", truncated);\n        }\n\n        if !cleaned.trim().is_empty() {\n            self.append_message(ConversationMessage::user(format!(\n                ""<stdout>\n{}\n</stdout>"",\n                cleaned\n            )));\n        }\n        self.terminal_output_buffer.clear();\n    }\n\n    /// Flush pending edits for a specific file.\n    pub fn flush_pending_edit_for_file(&mut self, target_file: &str) {\n        let before_snapshot = match self.pending_edits_before.get(target_file) {\n            Some(Some(s)) => s.clone(),\n            _ => return,\n        };\n\n        let after_state = self.file_states.get(target_file).cloned().unwrap_or_default();\n\n        if before_snapshot.trim_end_matches('\n') == after_state.trim_end_matches('\n') {\n            self.pending_edits_before.insert(target_file.to_string(), None);\n            self.pending_edit_regions.insert(target_file.to_string(), None);\n            return;\n        }\n\n        let changed = compute_changed_block_lines(&before_snapshot, &after_state)\n            .expect(""Failed to compute changed block lines"");\n\n        let before_total_lines = before_snapshot.split('\n').count();\n        let sed_cmd: String;\n\n        if changed.end_before < changed.start_before {\n            // Pure insertion\n            let escaped_lines: Vec<String> = changed\n                .replacement_lines\n                .iter()\n                .map(|line| escape_single_quotes_for_sed(line))\n                .collect();\n            let sed_payload = escaped_lines.join(""\n"");\n            if changed.start_before <= before_total_lines.max(1) {\n                sed_cmd = format!(\n                    ""sed -i '{}i\\\n{}' {}"",\n                    changed.start_before, sed_payload, target_file\n                );\n            } else {\n                sed_cmd = format!(""sed -i '$a\\\n{}' {}"", sed_payload, target_file);\n            }\n        } else if changed.replacement_lines.is_empty() {\n            // Pure deletion\n            sed_cmd = format!(\n                ""sed -i '{},{}d' {}"",\n                changed.start_before, changed.end_before, target_file\n            );\n        } else {\n            // Replacement\n            let escaped_lines: Vec<String> = changed\n                .replacement_lines\n                .iter()\n                .map(|line| escape_single_quotes_for_sed(line))\n                .collect();\n            let sed_payload = escaped_lines.join(""\n"");\n            sed_cmd = format!(\n                ""sed -i '{},{}c\\\n{}' {}"",\n                changed.start_before, changed.end_before, sed_payload, target_file\n            );\n        }\n\n        let total_lines = after_state.split('\n').count();\n        let center = (changed.start_after + changed.end_after) / 2;\n        let vp = serialize_compute_viewport(total_lines, center, self.config.viewport_radius);\n        self.per_file_viewport\n            .insert(target_file.to_string(), Some(vp));\n\n        self.maybe_capture_file_contents(target_file, &before_snapshot);\n\n        let chained_cmd = format!(\n            ""{} && cat -n {} | sed -n '{},{}p'"",\n            sed_cmd, target_file, vp.start, vp.end\n        );\n        self.append_message(ConversationMessage::assistant(fenced_block(\n            Some(""bash""),\n            &clean_text(&chained_cmd),\n        )));\n\n        let viewport_output = line_numbered_output(&after_state, Some(vp.start), Some(vp.end));\n        self.append_message(ConversationMessage::user(format!(\n            ""<stdout>\n{}\n</stdout>"",\n            viewport_output\n        )));\n\n        self.pending_edits_before.insert(target_file.to_string(), None);\n        self.pending_edit_regions.insert(target_file.to_string(), None);\n    }\n\n    /// Flush all pending edits.\n    pub fn flush_all_pending_edits(&mut self) {\n        let files: Vec<String> = self.pending_edits_before.keys().cloned().collect();\n        for file in files {\n            self.flush_pending_edit_for_file(&file);\n        }\n    }\n\n    /// Handle a tab (file switch) event.\n    pub fn handle_tab_event(&mut self, file_path: &str, text_content: Option<&str>) {\n        self.flush_all_pending_edits();\n        self.flush_terminal_output_buffer();\n\n        if let Some(text) = text_content {\n            let content = text.replace(""\\n"", ""\n"").replace(""\\r"", ""\r"");\n            self.file_states.insert(file_path.to_string(), content.clone());\n\n            let cmd = format!(""cat -n {}"", file_path);\n            self.append_message(ConversationMessage::assistant(fenced_block(\n                Some(""bash""),\n                &clean_text(&cmd),\n            )));\n            let output = line_numbered_output(&content, None, None);\n            self.append_message(ConversationMessage::user(format!(\n                ""<stdout>\n{}\n</stdout>"",\n                output\n            )));\n            self.files_opened_in_conversation.insert(file_path.to_string());\n        } else {\n            // File switch without content snapshot: show current viewport only\n            let content = self.file_states.get(file_path).cloned().unwrap_or_default();\n            let total_lines = content.split('\n').count();\n            let vp = self\n                .per_file_viewport\n                .get(file_path)\n                .and_then(|v| *v)\n                .filter(|v| v.end > 0)\n                .unwrap_or_else(|| {\n                    let new_vp = serialize_compute_viewport(total_lines, 1, self.config.viewport_radius);\n                    self.per_file_viewport.insert(file_path.to_string(), Some(new_vp));\n                    new_vp\n                });\n\n            if vp.end >= vp.start {\n                self.maybe_capture_file_contents(file_path, &content);\n                let cmd = format!(""cat -n {} | sed -n '{},{}p'"", file_path, vp.start, vp.end);\n                self.append_message(ConversationMessage::assistant(fenced_block(\n                    Some(""bash""),\n                    &clean_text(&cmd),\n                )));\n                let viewport_output = line_numbered_output(&content, Some(vp.start), Some(vp.end));\n                self.append_message(ConversationMessage::user(format!(\n                    ""<stdout>\n{}\n</stdout>"",\n                    viewport_output\n                )));\n            }\n        }\n    }\n\n    /// Handle a content change event.\n    pub fn handle_content_event(\n        &mut self,\n        file_path: &str,\n        offset: usize,\n        length: usize,\n        new_text: &str,\n    ) {\n        self.flush_terminal_output_buffer();\n\n        let before = self.file_states.get(file_path).cloned().unwrap_or_default();\n        let new_text_str = new_text;\n\n        // Approximate current edit region in line space\n        let start_line_current = before[..offset.min(before.len())].matches('\n').count() + 1;\n        let deleted_content = &before[offset.min(before.len())..(offset + length).min(before.len())];\n        let lines_added = new_text_str.matches('\n').count();\n        let lines_deleted = deleted_content.matches('\n').count();\n        let region_start = start_line_current;\n        let region_end = start_line_current + lines_added.max(lines_deleted);\n\n        // Flush pending edits if this edit is far from the pending region\n        let current_region = self.pending_edit_regions.get(file_path).and_then(|r| *r);\n        if let Some(region) = current_region {\n            if region_start < region.start.saturating_sub(self.config.coalesce_radius)\n                || region_start > region.end + self.config.coalesce_radius\n            {\n                self.flush_pending_edit_for_file(file_path);\n            }\n        }\n\n        let after = crate::helpers::apply_change(&before, offset, length, new_text);\n\n        if self.pending_edits_before.get(file_path).and_then(|v| v.as_ref()).is_none() {\n            self.pending_edits_before\n                .insert(file_path.to_string(), Some(before));\n        }\n\n        // Update/initialize region union\n        let current_region = self.pending_edit_regions.get(file_path).and_then(|r| *r);\n        let new_region = if let Some(region) = current_region {\n            EditRegion {\n                start: region.start.min(region_start),\n                end: region.end.max(region_end),\n            }\n        } else {\n            EditRegion {\n                start: region_start,\n                end: region_start.max(region_end),\n            }\n        };\n        self.pending_edit_regions\n            .insert(file_path.to_string(), Some(new_region));\n\n        self.file_states.insert(file_path.to_string(), after);\n    }\n\n    /// Handle a selection event.\n    pub fn handle_selection_event(&mut self, file_path: &str, offset: usize) {\n        // During an edit burst (pending edits), suppress viewport emissions\n        if self.pending_edits_before.get(file_path).and_then(|v| v.as_ref()).is_some() {\n            return;\n        }\n\n        self.flush_terminal_output_buffer();\n\n        let content = self.file_states.get(file_path).cloned().unwrap_or_default();\n        let total_lines = content.split('\n').count();\n        let target_line = content[..offset.min(content.len())].matches('\n').count() + 1;\n\n        let current_vp = self.per_file_viewport.get(file_path).and_then(|v| *v);\n        let mut should_emit = false;\n\n        let vp = if let Some(vp) = current_vp.filter(|v| v.end > 0) {\n            if target_line < vp.start || target_line > vp.end {\n                let new_vp =\n                    serialize_compute_viewport(total_lines, target_line, self.config.viewport_radius);\n                self.per_file_viewport\n                    .insert(file_path.to_string(), Some(new_vp));\n                should_emit = true;\n                new_vp\n            } else {\n                vp\n            }\n        } else {\n            let new_vp =\n                serialize_compute_viewport(total_lines, target_line, self.config.viewport_radius);\n            self.per_file_viewport\n                .insert(file_path.to_string(), Some(new_vp));\n            should_emit = true;\n            new_vp\n        };\n\n        if should_emit && vp.end >= vp.start {\n            self.maybe_capture_file_contents(file_path, &content);\n            let cmd = format!(""cat -n {} | sed -n '{},{}p'"", file_path, vp.start, vp.end);\n            self.append_message(ConversationMessage::assistant(fenced_block(\n                Some(""bash""),\n                &clean_text(&cmd),\n            )));\n            let viewport_output = line_numbered_output(&content, Some(vp.start), Some(vp.end));\n            self.append_message(ConversationMessage::user(format!(\n                ""<stdout>\n{}\n</stdout>"",\n                viewport_output\n            )));\n        }\n    }\n\n    /// Handle a terminal command event.\n    pub fn handle_terminal_command_event(&mut self, command: &str) {\n        self.flush_all_pending_edits();\n        self.flush_terminal_output_buffer();\n\n        let command_str = command.replace(""\\n"", ""\n"").replace(""\\r"", ""\r"");\n        self.append_message(ConversationMessage::assistant(fenced_block(\n            Some(""bash""),\n            &clean_text(&command_str),\n        )));\n    }\n\n    /// Handle a terminal output event.\n    pub fn handle_terminal_output_event(&mut self, output: &str) {\n        let raw_output = output.replace(""\\n"", ""\n"").replace(""\\r"", ""\r"");\n        self.terminal_output_buffer.push(raw_output);\n    }\n\n    /// Handle a terminal focus event.\n    pub fn handle_terminal_focus_event(&mut self) {\n        self.flush_all_pending_edits();\n        self.flush_terminal_output_buffer();\n        // No-op for bash transcript; focus changes don't emit commands/output\n    }\n\n    /// Handle a git branch checkout event.\n    pub fn handle_git_branch_checkout_event(&mut self, branch_info: &str) {\n        self.flush_all_pending_edits();\n        self.flush_terminal_output_buffer();\n\n        let branch_str = branch_info.replace(""\\n"", ""\n"").replace(""\\r"", ""\r"");\n        let cleaned = clean_text(&branch_str);\n\n        // Extract branch name from ""to 'branch_name'"" pattern\n        let re = regex::Regex::new(r""to '([^']+)'"").unwrap();\n        let branch_name = match re.captures(&cleaned) {\n            Some(caps) => caps.get(1).map(|m| m.as_str().trim().to_string()),\n            None => {\n                eprintln!(\n                    ""[crowd-pilot] Could not extract branch name from git checkout message: {}"",\n                    cleaned\n                );\n                return;\n            }\n        };\n\n        let mut branch_name = match branch_name {\n            Some(b) => b,\n            None => return,\n        };\n\n        // Safe-quote branch if it contains special characters\n        let special_chars = regex::Regex::new(r""[^A-Za-z0-9._/\\-]"").unwrap();\n        if special_chars.is_match(&branch_name) {\n            branch_name = format!(""'{}'"", branch_name.replace('\'', ""'\""'\""'""));\n        }\n\n        let cmd = format!(""git checkout {}"", branch_name);\n        self.append_message(ConversationMessage::assistant(fenced_block(\n            Some(""bash""),\n            &clean_text(&cmd),\n        )));\n    }\n\n    /// Finalize and get conversation ready for model.\n    pub fn finalize_for_model(&mut self) -> Vec<ConversationMessage> {\n        self.flush_all_pending_edits();\n        self.flush_terminal_output_buffer();\n        self.get_messages()\n    }\n}\n\n#[cfg(test)]\nmod tests {\n    use super::*;\n\n    /// Character-based approximate tokenizer for tests.\n    struct CharApproxTokenizer;\n\n    impl Tokenizer for CharApproxTokenizer {\n        fn count_tokens(&self, text: &str) -> usize {\n            text.len() / 4\n        }\n\n        fn truncate_to_max_tokens(&self, text: &str, max_tokens: usize) -> String {\n            text.chars().take(max_tokens * 4).collect()\n        }\n    }\n\n    #[test]\n    fn test_basic_tab_event() {\n        let mut manager =\n            ConversationStateManager::new(CharApproxTokenizer, ConversationStateManagerConfig::default());\n\n        manager.handle_tab_event(""/test/file.rs"", Some(""fn main() {\n    println!(\""hello\"");\n}""));\n\n        let messages = manager.finalize_for_model();\n        assert_eq!(messages.len(), 2);\n        assert_eq!(messages[0].from, ""Assistant"");\n        assert!(messages[0].value.contains(""cat -n /test/file.rs""));\n        assert_eq!(messages[1].from, ""User"");\n        assert!(messages[1].value.contains(""<stdout>""));\n    }\n\n    #[test]\n    fn test_content_event() {\n        let mut manager =\n            ConversationStateManager::new(CharApproxTokenizer, ConversationStateManagerConfig::default());\n\n        manager.handle_tab_event(""/test/file.rs"", Some(""line1\nline2\nline3""));\n        manager.handle_content_event(""/test/file.rs"", 6, 5, ""modified"");\n\n        let messages = manager.finalize_for_model();\n        // Should have: cat (open file), stdout, sed (edit), stdout\n        assert!(messages.len() >= 4);\n    }\n\n    #[test]\n    fn test_terminal_command() {\n        let mut manager =\n            ConversationStateManager::new(CharApproxTokenizer, ConversationStateManagerConfig::default());\n\n        manager.handle_terminal_command_event(""cargo build"");\n        manager.handle_terminal_output_event(""Compiling...\n"");\n        manager.handle_terminal_output_event(""Finished\n"");\n\n        let messages = manager.finalize_for_model();\n        assert_eq!(messages.len(), 2);\n        assert!(messages[0].value.contains(""cargo build""));\n        assert!(messages[1].value.contains(""Compiling""));\n    }\n}\n\n",rust,tab
+2,2044,"crates/core/src/conversation.rs",13552,0,"",rust,selection_command
+3,2156,"crates/core/src/conversation.rs",13520,0,"",rust,selection_command
+4,3714,"crates/core/src/conversation.rs",13458,0,"",rust,selection_command
+5,3851,"crates/core/src/conversation.rs",13441,0,"",rust,selection_command
+6,3865,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"5:48:18 PM [info] Activating crowd-code\n5:48:18 PM [info] Recording started\n5:48:18 PM [info] Initializing git provider using file system watchers...\n5:48:20 PM [info] Retrying git provider initialization...\n",Log,tab
+7,4241,"extension-output-pdoom-org.crowd-code-#1-crowd-code",208,0,"5:48:22 PM [info] Git repository found\n5:48:22 PM [info] Git provider initialized successfully\n5:48:22 PM [info] Initial git state: [object Object]\n5:48:22 PM [info] Git repository found\n5:48:22 PM [info] Git provider initialized successfully\n5:48:22 PM [info] Initial git state: [object Object]\n",Log,content
+8,9926,"crates/core/src/conversation.rs",0,0,"",rust,tab
+9,10302,"crates/core/src/conversation.rs",13418,0,"",rust,selection_command
+10,10440,"crates/core/src/conversation.rs",13441,0,"",rust,selection_command
+11,10607,"crates/core/src/conversation.rs",13458,0,"",rust,selection_command
+12,11952,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+13,12962,"TERMINAL",0,0,"",,terminal_focus
+14,12963,"crates/core/src/conversation.rs",0,0,"",rust,tab
+15,20574,"TERMINAL",0,0,"squeue",,terminal_command
+16,20584,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  36313   xiao.liu interacti     1   128  R 2025-12-13T12:56:48 2025-12-13T12:56:48  4:51:51   23:59:00 hai005\r\n  36303   xiao.liu interacti     1   128  R 2025-12-12T23:25:52 2025-12-12T23:25:52 18:22:47   23:59:00 hai006\r\n  36318 mihir.maha  standard     1    10  R 2025-12-13T15:53:27 2025-12-13T16:02:47  1:45:52 1-00:00:00 hai008\r\n  36317 mihir.maha  standard     1    10  R 2025-12-13T15:27:10 2025-12-13T15:27:10  2:21:29 1-00:00:00 hai004\r\n  36314   xiao.liu  standard     1   128  R 2025-12-13T13:27:57 2025-12-13T15:13:56  2:34:43   23:59:00 hai007\r\n  36304 nishant.ku  standard     3   624  R 2025-12-13T07:56:43 2025-12-13T07:56:43  9:51:56 1-00:00:00 hai[001-003]\r\n]0;franz.srambical@hai-login2:~/crowd-pilot-serializer",,terminal_output
+17,146587,"TERMINAL",0,0,"salloc --gpus=1 --ntasks-per-node=1 --cpus-per-task=10 --mem=400G --qos=normal",,terminal_command
+18,146631,"TERMINAL",0,0,"]633;Csalloc: Granted job allocation 36320\r\n",,terminal_output
+19,146703,"TERMINAL",0,0,"salloc: Nodes hai006 are ready for job\r\n",,terminal_output
+20,147161,"TERMINAL",0,0,"Running inside SLURM, Job ID 36320.\r\n]0;franz.srambical@hai-login2:~/crowd-pilot-serializer[?2004h[franz.srambical@hai006.haicore.berlin:~/crowd-pilot-serializer] $ ",,terminal_output
+21,221349,"crates/core/src/conversation.rs",13520,0,"",rust,selection_command
+22,221470,"crates/core/src/conversation.rs",13458,0,"",rust,selection_command
+23,275852,"TERMINAL",0,0,"\r(reverse-i-search)`': ",,terminal_output
+24,277201,"TERMINAL",0,0,"s': squeueo': rm -rf package-lock.json \ru': source /home/franz.srambical/crowd-pilot/.nemo_venv/bin/activate\r",,terminal_output
+25,277830,"TERMINAL",0,0,"[1@r': sour[1@c': sourc",,terminal_output
+26,278404,"TERMINAL",0,0,"[1@e': source",,terminal_output
+27,278627,"TERMINAL",0,0,"\r[39@[franz.srambical@hai006.haicore.berlin:~/crowd-pilot-serializer] $ source",,terminal_output
+28,281394,"TERMINAL",0,0,"",,terminal_output
+29,281518,"TERMINAL",0,0,"",,terminal_output
+30,281819,"TERMINAL",0,0,"",,terminal_output
+31,282928,"TERMINAL",0,0,"",,terminal_output
+32,282972,"TERMINAL",0,0,"",,terminal_output
+33,283569,"TERMINAL",0,0,"",,terminal_output
+34,284556,"TERMINAL",0,0,"",,terminal_output
+35,284809,"TERMINAL",0,0,"",,terminal_output
+36,285373,"TERMINAL",0,0,"",,terminal_output
+37,285680,"TERMINAL",0,0,"",,terminal_output
+38,286212,"TERMINAL",0,0,"",,terminal_output
+39,286755,"TERMINAL",0,0,"",,terminal_output
+40,287048,"TERMINAL",0,0,"",,terminal_output
+41,287162,"TERMINAL",0,0,"",,terminal_output
+42,287223,"TERMINAL",0,0,"",,terminal_output
+43,287663,"TERMINAL",0,0,"",,terminal_output
+44,288039,"TERMINAL",0,0,"",,terminal_output
+45,288661,"TERMINAL",0,0,"",,terminal_output
+46,289018,"TERMINAL",0,0,"",,terminal_output
+47,289173,"TERMINAL",0,0,"",,terminal_output
+48,289924,"TERMINAL",0,0,"",,terminal_output
+49,290199,"TERMINAL",0,0,"",,terminal_output
+50,290949,"TERMINAL",0,0,"",,terminal_output
+51,291069,"TERMINAL",0,0,"",,terminal_output
+52,291325,"TERMINAL",0,0,"",,terminal_output
+53,291575,"TERMINAL",0,0,"",,terminal_output
+54,292225,"TERMINAL",0,0,"",,terminal_output
+55,307548,"crates/core/src/conversation.rs",13520,0,"",rust,selection_command
+56,307682,"crates/core/src/conversation.rs",13458,0,"",rust,selection_command
+57,337815,"crates/core/src/conversation.rs",13441,0,"",rust,selection_command
+58,337922,"crates/core/src/conversation.rs",13458,0,"",rust,selection_command
+59,364265,"crates/cli/src/main.rs",0,0,"//! CLI tool for serializing crowd-pilot IDE interaction data.\n//!\n//! This tool processes CSV session files and outputs JSONL format suitable for\n//! NeMo SFT training. It uses an embedded Python interpreter to load HuggingFace\n//! tokenizers for accurate token counting.\n\nuse std::path::PathBuf;\n\nuse clap::Parser;\nuse pyo3::prelude::*;\nuse pyo3::types::PyModule;\n\nuse crowd_pilot_serializer_core::{\n    pipeline::{PipelineConfig, PipelineResult},\n    process_all_sessions, write_jsonl_output, Tokenizer,\n};\n\n/// Serialize crowd-pilot CSV sessions to NeMo JSONL format.\n#[derive(Parser, Debug)]\n#[command(name = ""crowd-pilot-serialize"")]\n#[command(author, version, about, long_about = None)]\nstruct Args {\n    /// Root directory containing CSV session files\n    #[arg(long)]\n    csv_root: PathBuf,\n\n    /// Output directory for JSONL files\n    #[arg(long)]\n    output_dir: PathBuf,\n\n    /// HuggingFace tokenizer model name or path\n    #[arg(long)]\n    tokenizer: String,\n\n    /// Maximum tokens per conversation chunk\n    #[arg(long, default_value = ""8192"")]\n    max_tokens_per_conversation: usize,\n\n    /// Maximum tokens per message\n    #[arg(long, default_value = ""2048"")]\n    max_tokens_per_message: usize,\n\n    /// Minimum messages required to keep a conversation\n    #[arg(long, default_value = ""5"")]\n    min_conversation_messages: usize,\n\n    /// Viewport radius (lines above/below cursor)\n    #[arg(long, default_value = ""10"")]\n    viewport_radius: usize,\n\n    /// Coalesce radius for grouping nearby edits\n    #[arg(long, default_value = ""5"")]\n    coalesce_radius: usize,\n\n    /// Fraction of sessions for validation (0.0-1.0)\n    #[arg(long, default_value = ""0.1"")]\n    val_ratio: f64,\n\n    /// Custom system prompt (optional)\n    #[arg(long)]\n    system_prompt: Option<String>,\n}\n\nconst DEFAULT_SYSTEM_PROMPT: &str = r#""You are a helpful assistant that can interact multiple times with a computer shell to solve programming tasks.\nYour response must contain exactly ONE bash code block with ONE command (or commands connected with && or ||).\n\nFormat your response as shown in <format_example>.\n\n<format_example>\n```bash\nyour_command_here\n```\n</format_example>\n\nFailure to follow these rules will cause your response to be rejected.""#;\n\n/// Wrapper around Python tokenizer for exact token counting and truncation.\nstruct PythonTokenizer {\n    tokenizer: Py<PyAny>,\n}\n\nimpl PythonTokenizer {\n    /// Load a HuggingFace tokenizer.\n    fn load(model_name: &str) -> PyResult<Self> {\n        Python::with_gil(|py| {\n            let transformers = PyModule::import(py, ""transformers"")?;\n            let auto_tokenizer = transformers.getattr(""AutoTokenizer"")?;\n            let tokenizer = auto_tokenizer.call_method1(""from_pretrained"", (model_name,))?;\n            Ok(Self {\n                tokenizer: tokenizer.into(),\n            })\n        })\n    }\n}\n\nimpl Tokenizer for PythonTokenizer {\n    fn count_tokens(&self, text: &str) -> usize {\n        Python::with_gil(|py| {\n            let tokenizer = self.tokenizer.as_ref(py);\n            let tokens = tokenizer\n                .call_method1(""encode"", (text,))\n                .expect(""Failed to encode text with tokenizer"");\n            tokens.len().unwrap()\n        })\n    }\n\n    fn truncate_to_max_tokens(&self, text: &str, max_tokens: usize) -> String {\n        Python::with_gil(|py| {\n            let tokenizer = self.tokenizer.as_ref(py);\n            let kwargs = pyo3::types::PyDict::new(py);\n            kwargs.set_item(""max_length"", max_tokens).unwrap();\n            kwargs.set_item(""truncation"", true).unwrap();\n            \n            let tokens = tokenizer\n                .call_method(""encode"", (text,), Some(kwargs))\n                .expect(""Failed to encode text with tokenizer"");\n            \n            tokenizer\n                .call_method1(""decode"", (tokens,))\n                .expect(""Failed to decode tokens"")\n                .extract()\n                .unwrap()\n        })\n    }\n}\n\nfn main() -> Result<(), Box<dyn std::error::Error>> {\n    let args = Args::parse();\n\n    println!(""Loading tokenizer from {}..."", args.tokenizer);\n    let tokenizer = PythonTokenizer::load(&args.tokenizer)?;\n\n    let config = PipelineConfig {\n        max_tokens_per_conversation: args.max_tokens_per_conversation,\n        max_tokens_per_message: args.max_tokens_per_message,\n        min_conversation_messages: args.min_conversation_messages,\n        viewport_radius: args.viewport_radius,\n        coalesce_radius: args.coalesce_radius,\n        val_ratio: args.val_ratio,\n    };\n\n    println!(""Processing CSV files from {:?}..."", args.csv_root);\n    let session_results = process_all_sessions(\n        &args.csv_root,\n        &tokenizer,\n        &config,\n    )?;\n\n    let total_sessions = session_results.len();\n    println!(""Processed {} sessions"", total_sessions);\n\n    let system_prompt = args.system_prompt.as_deref().unwrap_or(DEFAULT_SYSTEM_PROMPT);\n\n    println!(""Writing output to {:?}..."", args.output_dir);\n    let result: PipelineResult = write_jsonl_output(\n        session_results,\n        &args.output_dir,\n        args.val_ratio,\n        system_prompt,\n    )?;\n\n    let metadata_path = args.output_dir.join(""metadata.json"");\n    let metadata = serde_json::json!({\n        ""config"": {\n            ""csv_root"": args.csv_root.to_string_lossy(),\n            ""output_dir"": args.output_dir.to_string_lossy(),\n            ""tokenizer"": args.tokenizer,\n            ""max_tokens_per_conversation"": args.max_tokens_per_conversation,\n            ""max_tokens_per_message"": args.max_tokens_per_message,\n            ""min_conversation_messages"": args.min_conversation_messages,\n            ""viewport_radius"": args.viewport_radius,\n            ""coalesce_radius"": args.coalesce_radius,\n            ""val_ratio"": args.val_ratio,\n        },\n        ""counts"": {\n            ""total_sessions"": result.total_sessions,\n            ""total_conversations"": result.total_conversations,\n            ""train_conversations"": result.train_conversations,\n            ""val_conversations"": result.val_conversations,\n        },\n        ""stats"": {\n            ""total_messages"": result.total_messages,\n            ""total_tokens"": result.total_tokens,\n            ""avg_messages_per_conversation"": if result.total_conversations > 0 {\n                result.total_messages as f64 / result.total_conversations as f64\n            } else {\n                0.0\n            },\n            ""avg_tokens_per_conversation"": if result.total_conversations > 0 {\n                result.total_tokens as f64 / result.total_conversations as f64\n            } else {\n                0.0\n            },\n        },\n        ""files"": {\n            ""train_path"": args.output_dir.join(""training.jsonl"").to_string_lossy(),\n            ""val_path"": args.output_dir.join(""validation.jsonl"").to_string_lossy(),\n        },\n    });\n    std::fs::write(&metadata_path, serde_json::to_string_pretty(&metadata)?)?;\n\n    println!(""\n[summary]"");\n    println!(""  Total sessions processed: {}"", result.total_sessions);\n    println!(""  Train conversations: {}"", result.train_conversations);\n    println!(""  Val conversations: {}"", result.val_conversations);\n    println!(""  Total messages: {}"", result.total_messages);\n    println!(""  Total tokens: {}"", result.total_tokens);\n    println!(""  Output: {:?}/{{training,validation}}.jsonl"", args.output_dir);\n    println!(""  Metadata: {:?}"", metadata_path);\n\n    Ok(())\n}\n\n",rust,tab
+60,392539,"TERMINAL",0,0,"\r[franz.srambical@hai006.haicore.berlin:~/crowd-pilot-serializer] $ ",,terminal_output
+61,439804,"TERMINAL",0,0,"m",,terminal_output
+62,440745,"TERMINAL",0,0,"kdir ",,terminal_output
+63,440745,"TERMINAL",0,0,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/",,terminal_output
+64,441743,"TERMINAL",0,0,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/",,terminal_output
+65,441743,"TERMINAL",0,0,"\r\ngenerate_array_record_dataset.sh               qwen-coder-small/\r\ngenerate_array_record_dataset_bash_version.sh  qwen_0_6/\r\ngenerate_parquet_dataset.sh                    random-weights/\r\nnemo/                                          tabbing_model/\r\nqwen-coder-large/                              \r\n[franz.srambical@hai006.haicore.berlin:~/crowd-pilot-serializer] $ mkdir /home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/",,terminal_output
+66,446048,"TERMINAL",0,0,"c",,terminal_output
+67,446463,"TERMINAL",0,0,"r",,terminal_output
+68,446663,"TERMINAL",0,0,"ow",,terminal_output
+69,446954,"TERMINAL",0,0,"d",,terminal_output
+70,447380,"TERMINAL",0,0,"-",,terminal_output
+71,447590,"TERMINAL",0,0,"",,terminal_output
+72,448743,"TERMINAL",0,0,"_p",,terminal_output
+73,448798,"TERMINAL",0,0,"i",,terminal_output
+74,448947,"TERMINAL",0,0,"o \r",,terminal_output
+75,449666,"TERMINAL",0,0,"\r\n\r",,terminal_output
+76,450585,"TERMINAL",0,0,"l \r",,terminal_output
+77,450862,"TERMINAL",0,0,"ot",,terminal_output
+78,451147,"TERMINAL",0,0,"_",,terminal_output
+79,451270,"TERMINAL",0,0,"se",,terminal_output
+80,451370,"TERMINAL",0,0,"r",,terminal_output
+81,452269,"TERMINAL",0,0,"zi",,terminal_output
+82,452534,"TERMINAL",0,0,"",,terminal_output
+83,452946,"TERMINAL",0,0,"iz",,terminal_output
+84,453393,"TERMINAL",0,0,"a",,terminal_output
+85,454446,"TERMINAL",0,0,"l",,terminal_output
+86,454783,"TERMINAL",0,0,"ize",,terminal_output
+87,454984,"TERMINAL",0,0,"r",,terminal_output
+88,455052,"TERMINAL",0,0,"\r\n[?2004l\r]0;franz.srambical@hai-login2:~/crowd-pilot-serializer[?2004h[franz.srambical@hai006.haicore.berlin:~/crowd-pilot-serializer] $ ",,terminal_output
+89,457361,"TERMINAL",0,0,"c",,terminal_output
+90,457444,"TERMINAL",0,0,"d",,terminal_output
+91,457828,"TERMINAL",0,0," ",,terminal_output
+92,460688,"TERMINAL",0,0,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer",,terminal_output
+93,461044,"TERMINAL",0,0,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer\r\n[?2004l\r]0;franz.srambical@hai-login2:~/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer[?2004h[franz.srambical@hai006.haicore.berlin:~/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer] $ ",,terminal_output
+94,462574,"TERMINAL",0,0,"t",,terminal_output
+95,462717,"TERMINAL",0,0,"o",,terminal_output
+96,462972,"TERMINAL",0,0,"uc",,terminal_output
+97,463143,"TERMINAL",0,0,"h",,terminal_output
+98,466651,"TERMINAL",0,0," ",,terminal_output
+99,467527,"TERMINAL",0,0,"ser",,terminal_output
+100,467631,"TERMINAL",0,0,"i",,terminal_output
+101,467923,"TERMINAL",0,0,"a",,terminal_output
+102,468119,"TERMINAL",0,0,"li",,terminal_output
+103,468242,"TERMINAL",0,0,"ze",,terminal_output
+104,468831,"TERMINAL",0,0,".sh",,terminal_output
+105,470651,"TERMINAL",0,0,"\r\n[?2004l\r]0;franz.srambical@hai-login2:~/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer[?2004h[franz.srambical@hai006.haicore.berlin:~/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer] $ ",,terminal_output
+106,478951,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize.sh",0,0,"",shellscript,tab
+107,487249,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize.sh",0,0,"cargo run --bin crowd-pilot-serialize -- --help 2>&1 | cat",shellscript,content
+108,487250,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize.sh",58,0,"",shellscript,selection_keyboard
+109,487808,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize.sh",57,0,"",shellscript,selection_command
+110,487864,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize.sh",55,0,"",shellscript,selection_command
+111,488070,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize.sh",53,0,"",shellscript,selection_command
+112,488177,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize.sh",51,0,"",shellscript,selection_command
+113,488434,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize.sh",49,0,"",shellscript,selection_command
+114,488521,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize.sh",48,0,"",shellscript,selection_command
+115,488949,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize.sh",43,0,"",shellscript,selection_command
+116,489159,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize.sh",41,0,"",shellscript,selection_command
+117,489377,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize.sh",38,0,"",shellscript,selection_command
+118,489957,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize.sh",37,0,"",shellscript,selection_command
+119,492780,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize.sh",37,21,"",shellscript,content
+120,493748,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize.sh",36,0,"",shellscript,selection_command
+121,496633,"TERMINAL",0,0,"car",,terminal_output
+122,496688,"TERMINAL",0,0,"go",,terminal_output
+123,496743,"TERMINAL",0,0," r",,terminal_output
+124,497459,"TERMINAL",0,0,"un",,terminal_output
+125,497787,"TERMINAL",0,0,"",,terminal_output
+126,498368,"TERMINAL",0,0,"c",,terminal_output
+127,498675,"TERMINAL",0,0,"d\r\n[?2004l\r]0;franz.srambical@hai-login2:~[?2004h[franz.srambical@hai006.haicore.berlin:~] $ c",,terminal_output
+128,498724,"TERMINAL",0,0,"d ",,terminal_output
+129,500238,"TERMINAL",0,0,"c",,terminal_output
+130,500329,"TERMINAL",0,0,"rowd",,terminal_output
+131,500677,"TERMINAL",0,0,"-",,terminal_output
+132,500677,"TERMINAL",0,0,"pi",,terminal_output
+133,500839,"TERMINAL",0,0,"lot",,terminal_output
+134,501737,"TERMINAL",0,0,"-",,terminal_output
+135,502926,"TERMINAL",0,0,"serializer/",,terminal_output
+136,503432,"TERMINAL",0,0,"\r\n[?2004l\r]0;franz.srambical@hai-login2:~/crowd-pilot-serializer[?2004h[franz.srambical@hai006.haicore.berlin:~/crowd-pilot-serializer] $ ",,terminal_output
+137,504618,"TERMINAL",0,0,"s",,terminal_output
+138,504994,"TERMINAL",0,0,"",,terminal_output
+139,505750,"TERMINAL",0,0,"ar",,terminal_output
+140,506068,"TERMINAL",0,0,"go",,terminal_output
+141,506236,"TERMINAL",0,0," run",,terminal_output
+142,506457,"TERMINAL",0,0," ",,terminal_output
+143,506644,"TERMINAL",0,0,"--",,terminal_output
+144,507080,"TERMINAL",0,0,"bin",,terminal_output
+145,507226,"TERMINAL",0,0,"",,terminal_output
+146,507382,"TERMINAL",0,0,"c",,terminal_output
+147,507836,"TERMINAL",0,0,"r",,terminal_output
+148,507836,"TERMINAL",0,0,"ow",,terminal_output
+149,508006,"TERMINAL",0,0,"d",,terminal_output
+150,508211,"TERMINAL",0,0,"-",,terminal_output
+151,508482,"TERMINAL",0,0,"pi",,terminal_output
+152,508724,"TERMINAL",0,0,"l",,terminal_output
+153,508859,"TERMINAL",0,0,"ot",,terminal_output
+154,508928,"TERMINAL",0,0,"-",,terminal_output
+155,509378,"TERMINAL",0,0,"ser",,terminal_output
+156,510282,"TERMINAL",0,0,"",,terminal_output
+157,510543,"TERMINAL",0,0,"",,terminal_output
+158,512021,"TERMINAL",0,0,"",,terminal_output
+159,513201,"TERMINAL",0,0,"",,terminal_output
+160,513643,"TERMINAL",0,0," crowd-pilot-ser",,terminal_output
+161,514665,"TERMINAL",0,0,"",,terminal_output
+162,515726,"TERMINAL",0,0,"i",,terminal_output
+163,515726,"TERMINAL",0,0,"a",,terminal_output
+164,515877,"TERMINAL",0,0,"lize",,terminal_output
+165,517390,"TERMINAL",0,0,"",,terminal_output
+166,517929,"TERMINAL",0,0,"[1@c",,terminal_output
+167,518157,"TERMINAL",0,0,"",,terminal_output
+168,518662,"TERMINAL",0,0,"\r\n[?2004l\r",,terminal_output
+169,522751,"TERMINAL",0,0,"   Compiling pyo3-build-config v0.20.3\r\n   Compiling crowd-pilot-serializer-core v0.1.0 (/fast/home/franz.srambical/crowd-pilot-serializer/crates/core)\r\n    Building [======================>    ] 72/84: pyo3-build-config(build), crowd-pilot-serializer-core                                  \r",,terminal_output
+170,522983,"TERMINAL",0,0,"    Building [======================>    ] 73/84: crowd-pilot-serializer-core, pyo3-build-config                                         \r",,terminal_output
+171,526114,"TERMINAL",0,0,"   Compiling pyo3-macros-backend v0.20.3\r\n    Building [======================>    ] 73/84: crowd-pilot-serializer-core, pyo3-macros-backend, pyo3-build-config                    \r",,terminal_output
+172,526336,"TERMINAL",0,0,"    Building [======================>    ] 74/84: pyo3-macros-backend, pyo3-build-config                                                 \r",,terminal_output
+173,526510,"TERMINAL",0,0,"   Compiling pyo3-ffi v0.20.3\r\n   Compiling pyo3 v0.20.3\r\n    Building [=======================>   ] 75/84: pyo3(build.rs), pyo3-ffi(build.rs), pyo3-macros-backend                                \r",,terminal_output
+174,528048,"TERMINAL",0,0,"   Compiling pyo3-macros v0.20.3\r\n    Building [=======================>   ] 76/84: pyo3(build.rs), pyo3-macros, pyo3-ffi(build.rs)                                        \r",,terminal_output
+175,529035,"TERMINAL",0,0,"    Building [=======================>   ] 77/84: pyo3(build.rs), pyo3-macros, pyo3-ffi(build)                                           \r    Building [========================>  ] 78/84: pyo3-macros, pyo3-ffi(build)                                                           \r    Building [========================>  ] 79/84: pyo3-ffi(build)                                                                        \r    Building [========================>  ] 80/84: pyo3(build), pyo3-ffi                                                                  \r    Building [=========================> ] 81/84: pyo3-ffi                                                                               \r",,terminal_output
+176,529180,"TERMINAL",0,0,"    Building [=========================> ] 81/84: pyo3-ffi, pyo3                                                                         \r",,terminal_output
+177,529240,"TERMINAL",0,0,"    Building [=========================> ] 82/84: pyo3                                                                                   \r",,terminal_output
+178,532262,"TERMINAL",0,0,"   Compiling crowd-pilot-serialize v0.1.0 (/fast/home/franz.srambical/crowd-pilot-serializer/crates/cli)\r\n    Building [=========================> ] 83/84: crowd-pilot-serialize(bin)                                                             \r",,terminal_output
+179,534691,"TERMINAL",0,0,"    Finished ]8;;https://doc.rust-lang.org/cargo/reference/profiles.html#default-profiles\`dev` profile [unoptimized + debuginfo]]8;;\ target(s) in 15.33s\r\n     Running `target/debug/crowd-pilot-serialize`\r\ntarget/debug/crowd-pilot-serialize: error while loading shared libraries: libpython3.12.so.1.0: cannot open shared object file: No such file or directory\r\n]0;franz.srambical@hai-login2:~/crowd-pilot-serializer[?2004h[franz.srambical@hai006.haicore.berlin:~/crowd-pilot-serializer] $ ",,terminal_output
+180,541235,"crates/cli/src/main.rs",0,0,"",rust,tab
+181,556263,"TERMINAL",0,0,"cargo run --bin crowd-pilot-serialize",,terminal_output
+182,556333,"TERMINAL",0,0," ",,terminal_output
+183,556637,"TERMINAL",0,0,"-",,terminal_output
+184,557201,"TERMINAL",0,0,"-",,terminal_output
+185,557482,"TERMINAL",0,0,"help",,terminal_output
+186,557772,"TERMINAL",0,0,"\r\n[?2004l\rRun a binary or example of the local package\r\n\r\nUsage: cargo run [OPTIONS] [ARGS]...\r\n\r\nArguments:\r\n  [ARGS]...  Arguments for the binary or example to run\r\n\r\nOptions:\r\n      --message-format <FMT>     Error format [possible values: human, short, json, json-diagnostic-short, json-diagnostic-rendered-ansi,\r\n                                 json-render-diagnostics]\r\n  -v, --verbose...               Use verbose output (-vv very verbose/build.rs output)\r\n  -q, --quiet                    Do not print cargo log messages\r\n      --color <WHEN>             Coloring [possible values: auto, always, never]\r\n      --config <KEY=VALUE|PATH>  Override a configuration value\r\n  -Z <FLAG>                      Unstable (nightly-only) flags to Cargo, see 'cargo -Z help' for details\r\n  -h, --help                     Print help\r\n\r\nPackage Selection:\r\n  -p, --package [<SPEC>]  Package with the target to run\r\n\r\nTarget Selection:\r\n      --bin [<NAME>]      Name of the bin target to run\r\n      --example [<NAME>]  Name of the example target to run\r\n\r\nFeature Selection:\r\n  -F, --features <FEATURES>  Space or comma separated list of features to activate\r\n      --all-features         Activate all available features\r\n      --no-default-features  Do not activate the `default` feature\r\n\r\nCompilation Options:\r\n  -j, --jobs <N>                Number of parallel jobs, defaults to # of CPUs.\r\n      --keep-going              Do not abort the build as soon as there is an error\r\n  -r, --release                 Build artifacts in release mode, with optimizations\r\n      --profile <PROFILE-NAME>  Build artifacts with the specified profile\r\n      --target [<TRIPLE>]       Build for the target triple\r\n      --target-dir <DIRECTORY>  Directory for all generated artifacts\r\n      --unit-graph              Output build graph in JSON (unstable)\r\n      --timings[=<FMTS>]        Timing output formats (unstable) (comma separated): html, json\r\n\r\nManifest Options:\r\n      --manifest-path <PATH>  Path to Cargo.toml\r\n      --lockfile-path <PATH>  Path to Cargo.lock (unstable)\r\n      --ignore-rust-version   Ignore `rust-version` specification in packages\r\n      --locked                Assert that `Cargo.lock` will remain unchanged\r\n      --offline               Run without accessing the network\r\n      --frozen                Equivalent to specifying both --locked and --offline\r\n\r\nRun `cargo help run` for more detailed information.\r\n]0;franz.srambical@hai-login2:~/crowd-pilot-serializer[?2004h[franz.srambical@hai006.haicore.berlin:~/crowd-pilot-serializer] $ ",,terminal_output
+187,567433,"TERMINAL",0,0,"cargo run --bin crowd-pilot-serialize --help",,terminal_output
+188,586016,"TERMINAL",0,0,"",,terminal_output
+189,586737,"TERMINAL",0,0,"",,terminal_output
+190,587196,"TERMINAL",0,0,"",,terminal_output
+191,588413,"TERMINAL",0,0,"[1@-",,terminal_output
+192,589018,"TERMINAL",0,0,"",,terminal_output
+193,589225,"TERMINAL",0,0," --help",,terminal_output
+194,589438,"TERMINAL",0,0,"- --help",,terminal_output
+195,589516,"TERMINAL",0,0,"- --help",,terminal_output
+196,592484,"TERMINAL",0,0,"",,terminal_output
+197,593107,"TERMINAL",0,0,"\r\n[?2004l\r",,terminal_output
+198,593498,"TERMINAL",0,0,"    Finished ]8;;https://doc.rust-lang.org/cargo/reference/profiles.html#default-profiles\`dev` profile [unoptimized + debuginfo]]8;;\ target(s) in 0.47s\r\n     Running `target/debug/crowd-pilot-serialize --help`\r\ntarget/debug/crowd-pilot-serialize: error while loading shared libraries: libpython3.12.so.1.0: cannot open shared object file: No such file or directory\r\n]0;franz.srambical@hai-login2:~/crowd-pilot-serializer[?2004h[franz.srambical@hai006.haicore.berlin:~/crowd-pilot-serializer] $ ",,terminal_output

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-7c4b7a0c-d468-4945-92ec-e9dc081250161766827133150-2025_12_27-10.22.14.278/source.csv

@@ -0,0 +1,58 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,1465,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"10:22:14 AM [info] Activating crowd-code\n10:22:14 AM [info] Recording started\n10:22:14 AM [info] Initializing git provider using file system watchers...\n10:22:15 AM [info] Git repository found\n10:22:15 AM [info] Git provider initialized successfully\n10:22:15 AM [info] Initial git state: [object Object]\n",Log,tab
+3,11424,"extension-output-pdoom-org.crowd-code-#1-crowd-code",304,0,"",Log,selection_mouse
+4,22228,"train_sft.py",0,0,"#!/usr/bin/env python3\n""""""\nRay-free SFT Training Script for Miles.\n\nThis script provides a simplified training path for Supervised Fine-Tuning (SFT)\nthat bypasses Ray entirely and uses torchrun for distributed training.\n\nUsage:\n    torchrun --nproc_per_node=2 train_sft.py \\n        --hf-checkpoint /path/to/model \\n        --prompt-data /path/to/data.parquet \\n        --input-key messages \\n        --apply-chat-template \\n        ...\n\nThis is equivalent to the Ray-based SFT with --debug-train-only, but without\nthe Ray overhead.\n""""""\n\nimport logging\nimport os\nfrom argparse import Namespace\nfrom datetime import timedelta\nfrom itertools import accumulate\n\nimport torch\nimport torch.distributed as dist\nfrom torch.distributed.device_mesh import init_device_mesh\nfrom tqdm import tqdm\nfrom transformers import AutoConfig\n\nfrom ring_flash_attn import substitute_hf_flash_attn, update_ring_flash_attn_params\n\nfrom miles.models.peft import LoRAConfig, apply_lora\nfrom miles.backends.fsdp_utils import checkpoint\nfrom miles.backends.fsdp_utils.actor import (\n    apply_fsdp2,\n    get_logprob_and_entropy_with_cp,\n    sum_of_sample_mean,\n)\nfrom miles.backends.fsdp_utils.data_packing import (\n    pack_sequences,\n    pad_packed_sequence_with_cp,\n    unpack_sequences,\n)\nfrom miles.backends.fsdp_utils.lr_scheduler import get_lr_scheduler\nfrom miles.rollout.data_source import RolloutDataSource\nfrom miles.utils import tracking_utils\nfrom miles.utils.arguments import parse_args\nfrom miles.utils.data import get_minimum_num_micro_batch_size\nfrom miles.utils.distributed_utils import get_gloo_group, init_gloo_group\nfrom miles.utils.logging_utils import configure_logger\nfrom miles.utils.mask_utils import MultiTurnLossMaskGenerator\nfrom miles.utils.misc import should_run_periodic_action\nfrom miles.utils.processing_utils import load_processor, load_tokenizer\nfrom miles.utils.profile_utils import TrainProfiler\nfrom miles.utils.timer import timer\nfrom miles.utils.tracking_utils import init_tracking\nfrom miles.utils.types import Sample\n\nlogger = logging.getLogger(__name__)\n\n\nclass SFTTrainer:\n    """"""\n    A simplified trainer for SFT that runs without Ray.\n\n    This class combines the functionality of:\n    - FSDPTrainRayActor (model initialization, FSDP wrapping, training)\n    - RolloutManager (data loading via generate_rollout)\n    - The main training loop from train.py\n    """"""\n\n    def __init__(self, args: Namespace):\n        self.args = args\n        self.device = torch.device(""cuda"")\n\n        self._init_distributed()\n\n        self._setup_device_mesh()\n\n        torch.manual_seed(args.seed)\n\n        self._enable_true_on_policy_optimizations()\n\n        if dist.get_rank() == 0:\n            init_tracking(args, primary=True)\n\n        self._load_tokenizer_and_config()\n\n        self._init_data_source()\n\n        self._init_model()\n\n        self._init_optimizer()\n\n        self._load_checkpoint()\n\n        self.prof = TrainProfiler(args)\n        self.prof.on_init_end()\n\n        logger.info(f""[Rank {dist.get_rank()}] SFTTrainer initialized successfully"")\n\n    def _init_distributed(self):\n        """"""Initialize distributed training.""""""\n        # torchrun sets these environment variables\n        local_rank = int(os.environ.get(""LOCAL_RANK"", 0))\n        torch.cuda.set_device(f""cuda:{local_rank}"")\n\n        backend = self.args.distributed_backend\n        dist.init_process_group(\n            backend=backend,\n            timeout=timedelta(minutes=self.args.distributed_timeout_minutes),\n        )\n        init_gloo_group()\n\n        self.args.rank = dist.get_rank()\n        self.args.world_size = dist.get_world_size()\n\n        logger.info(\n            f""[Rank {self.args.rank}] Distributed initialized: ""\n            f""world_size={self.args.world_size}, local_rank={local_rank}""\n        )\n\n    def _setup_device_mesh(self):\n        """"""Setup device mesh for FSDP (no context parallelism for SFT).""""""\n        world_size = dist.get_world_size()\n        rank = dist.get_rank()\n\n        self.cp_size = self.args.context_parallel_size\n        self.dp_size = world_size // self.cp_size\n\n        self.mesh = init_device_mesh(\n            ""cuda"",\n            mesh_shape=(self.dp_size, self.cp_size),\n            mesh_dim_names=(""dp"", ""cp""),\n        )\n\n        self.dp_group = self.mesh.get_group(""dp"")\n        self.cp_group = self.mesh.get_group(""cp"")\n        self.dp_mesh = self.mesh[""dp""]\n\n        self.dp_rank = rank // self.cp_size\n        self.cp_rank = rank % self.cp_size\n\n        logger.info(\n            f""[Rank {rank}] Device mesh: dp_size={self.dp_size}, cp_size={self.cp_size}, ""\n            f""dp_rank={self.dp_rank}, cp_rank={self.cp_rank}""\n        )\n\n        # Setup Ring Flash Attention with CP group from mesh (only when cp_size > 1)\n        if self.cp_size > 1:\n            substitute_hf_flash_attn(self.cp_group, heads_k_stride=1)\n            logger.info(f""[Rank {rank}] CP initialized via device mesh"")\n\n    def _enable_true_on_policy_optimizations(self):\n        """"""Enable true on-policy optimizations or apply MoE patches.""""""\n        if self.args.true_on_policy_mode:\n            from sglang.srt.batch_invariant_ops import enable_batch_invariant_mode\n\n            from miles.backends.fsdp_utils.models.qwen3_moe import (\n                apply_true_on_policy_patch_for_qwen3_moe,\n            )\n\n            logger.info(""SFTTrainer: enabling batch_invariant_mode for true-on-policy"")\n            enable_batch_invariant_mode(\n                # In Qwen3, rope uses bmm; disabling makes it aligned\n                enable_bmm=False,\n            )\n\n            apply_true_on_policy_patch_for_qwen3_moe()\n        else:\n            from miles.backends.fsdp_utils.models.qwen3_moe_hf import apply_fsdp_moe_patch\n\n            apply_fsdp_moe_patch()\n\n    def _load_tokenizer_and_config(self):\n        """"""Load tokenizer and model config sequentially to avoid race conditions.""""""\n        for i in range(dist.get_world_size()):\n            if i == dist.get_rank():\n                self.hf_config = AutoConfig.from_pretrained(\n                    self.args.hf_checkpoint, trust_remote_code=True\n                )\n                self.tokenizer = load_tokenizer(\n                    self.args.hf_checkpoint, trust_remote_code=True\n                )\n                self.processor = None\n                if self.args.multimodal_keys:\n                    self.processor = load_processor(\n                        self.args.hf_checkpoint, trust_remote_code=True\n                    )\n            dist.barrier(group=get_gloo_group())\n\n        # Initialize loss mask generator for SFT\n        self.mask_generator = MultiTurnLossMaskGenerator(\n            self.tokenizer,\n            tokenizer_type=getattr(self.args, ""loss_mask_type"", None),\n        )\n\n    def _init_data_source(self):\n        """"""Initialize the data source for SFT training.""""""\n        self.data_source = RolloutDataSource(self.args, self.args.prompt_data)\n        self.val_data_source = None\n        if self.args.val_prompt_data is not None:\n            self.val_data_source = RolloutDataSource(self.args, self.args.val_prompt_data)\n\n        # Calculate num_rollout from dataset size\n        if self.args.num_rollout is None:\n            num_rollout_per_epoch = len(self.data_source.dataset) // self.args.rollout_batch_size\n            self.args.num_rollout = num_rollout_per_epoch * self.args.num_epoch\n            self.num_rollout_per_epoch = num_rollout_per_epoch\n        else:\n            self.num_rollout_per_epoch = None\n\n        if getattr(self.args, ""start_rollout_id"", None) is None:\n            self.args.start_rollout_id = 0\n\n        logger.info(\n            f""[Rank {dist.get_rank()}] Data source initialized: ""\n            f""dataset_size={len(self.data_source.dataset)}, ""\n            f""num_rollout={self.args.num_rollout}""\n        )\n\n    def _get_init_weight_context_manager(self):\n        """"""Get context manager for model initialization.""""""\n        from accelerate import init_empty_weights\n\n        use_meta_tensor = not self.hf_config.tie_word_embeddings\n\n        def cpu_init_weights():\n            return torch.device(""cpu"")\n\n        if use_meta_tensor:\n            return init_empty_weights if dist.get_rank() != 0 else cpu_init_weights\n        else:\n            return cpu_init_weights\n\n    def _fsdp2_load_full_state_dict(self, model, full_state, device_mesh, cpu_offload):\n        """"""Load full state dict into FSDP2 model with broadcast from rank 0.""""""\n        from torch.distributed.checkpoint.state_dict import (\n            StateDictOptions,\n            set_model_state_dict,\n        )\n\n        if dist.get_rank() == 0:\n            model = model.to(device=torch.cuda.current_device(), non_blocking=True)\n        else:\n            model = model.to_empty(device=torch.cuda.current_device())\n\n        is_cpu_offload = cpu_offload is not None\n        options = StateDictOptions(\n            full_state_dict=True, cpu_offload=is_cpu_offload, broadcast_from_rank0=True\n        )\n\n        set_model_state_dict(model, full_state, options=options)\n\n        for _name, buf in model.named_buffers():\n            dist.broadcast(buf, src=0)\n\n        if is_cpu_offload:\n            model.to(""cpu"", non_blocking=True)\n            for buf in model.buffers():\n                buf.data = buf.data.to(torch.cuda.current_device())\n\n        return model\n\n    def _get_model_cls(self):\n        """"""Get the appropriate model class based on config.""""""\n        if hasattr(self.hf_config, ""vision_config""):\n            from transformers import AutoModelForImageTextToText\n\n            return AutoModelForImageTextToText\n        else:\n            from transformers import AutoModelForCausalLM\n\n            return AutoModelForCausalLM\n\n    def _init_model(self):\n        """"""Initialize and wrap model with FSDP.""""""\n        self.fsdp_cpu_offload = getattr(self.args, ""fsdp_cpu_offload"", False)\n\n        init_context = self._get_init_weight_context_manager()\n\n        with init_context():\n            model = self._get_model_cls().from_pretrained(\n                self.args.hf_checkpoint,\n                trust_remote_code=True,\n                attn_implementation=self.args.attn_implementation,\n            )\n\n            if self.args.use_lora:\n                lora_config = LoRAConfig(\n                    lora_rank=self.args.lora_rank,\n                    lora_alpha=self.args.lora_alpha,\n                    lora_dropout=self.args.lora_dropout,\n                    target_modules=self.args.lora_target_modules,\n                )\n                model = apply_lora(model, lora_config)\n                logger.info(f""[Rank {dist.get_rank()}] Applied LoRA: {lora_config}"")\n\n        model.train()\n        full_state = model.state_dict()\n\n        model = apply_fsdp2(\n            model, mesh=self.dp_mesh, cpu_offload=self.fsdp_cpu_offload, args=self.args\n        )\n\n        model = self._fsdp2_load_full_state_dict(\n            model,\n            full_state,\n            self.dp_mesh,\n            cpu_offload=True if self.fsdp_cpu_offload else None,\n        )\n\n        self.model = model\n\n        if self.args.gradient_checkpointing:\n            # FIXME: Conceptually, gradient checkpointing should be compatible with LoRA, but we don't support it yet.\n            assert not self.args.use_lora, ""Gradient checkpointing is incompatible with LoRA""\n            self.model.gradient_checkpointing_enable()\n\n        logger.info(f""[Rank {dist.get_rank()}] Model initialized with FSDP"")\n\n    def _init_optimizer(self):\n        """"""Initialize optimizer and learning rate scheduler.""""""\n        trainable_params = [p for p in self.model.parameters() if p.requires_grad]\n        \n        if self.args.use_lora:\n            total_params = sum(p.numel() for p in self.model.parameters())\n            trainable_count = sum(p.numel() for p in trainable_params)\n            logger.info(\n                f""[Rank {dist.get_rank()}] LoRA: {trainable_count:,} trainable params ""\n                f""out of {total_params:,} total ({100 * trainable_count / total_params:.2f}%)""\n            )\n        \n        if self.args.optimizer == ""adam"":\n            self.optimizer = torch.optim.AdamW(\n                trainable_params,\n                lr=self.args.lr,\n                betas=(self.args.adam_beta1, self.args.adam_beta2),\n                eps=self.args.adam_eps,\n                weight_decay=self.args.weight_decay,\n            )\n        else:\n            raise ValueError(f""Unsupported optimizer: {self.args.optimizer}"")\n\n        self.lr_scheduler = get_lr_scheduler(self.args, self.optimizer)\n        self.global_step = 0\n        self.micro_step = 0\n\n    def _load_checkpoint(self):\n        """"""Load checkpoint if available.""""""\n        checkpoint_payload = checkpoint.load(self)\n        checkpoint.finalize_load(self, checkpoint_payload)\n        \n        if self.args.rollout_global_dataset and self.args.start_rollout_id > 0:\n            self.data_source.load(self.args.start_rollout_id - 1)\n\n    def generate_sft_rollout(self, rollout_id: int, data_source: RolloutDataSource) -> list[Sample]:\n        """"""Generate SFT rollout data (tokenize and create loss masks).""""""\n        samples = data_source.get_samples(self.args.rollout_batch_size)\n\n        result = []\n        for i, (sample,) in enumerate(samples):\n            messages = sample.prompt\n            token_ids, loss_mask = self.mask_generator.get_loss_mask(messages)\n            response_length = self.mask_generator.get_response_lengths([loss_mask])[0]\n\n            sample.tokens = token_ids\n            sample.response_length = response_length\n            sample.reward = 0\n            sample.loss_mask = loss_mask[-response_length:]\n            result.append(sample)\n\n            if i == 0 and rollout_id == 0 and dist.get_rank() == 0:\n                logger.info(\n                    f""SFT rollout sample: tokens_len={len(token_ids)}, ""\n                    f""response_length={response_length}""\n                )\n\n        return result\n\n    def _convert_samples_to_train_data(self, samples: list[Sample]) -> dict:\n        """"""Convert samples to training data format.""""""\n        train_data = {\n            ""tokens"": [sample.tokens for sample in samples],\n            ""response_lengths"": [sample.response_length for sample in samples],\n            ""rewards"": [0.0 for _ in samples],\n            ""raw_reward"": [0.0 for _ in samples],\n            ""truncated"": [0 for _ in samples],\n            ""sample_indices"": [sample.index for sample in samples],\n        }\n\n        loss_masks = []\n        for sample in samples:\n            if sample.loss_mask is None:\n                sample.loss_mask = [1] * sample.response_length\n            loss_masks.append(sample.loss_mask)\n        train_data[""loss_masks""] = loss_masks\n\n        return train_data\n\n    def _split_train_data_by_dp(self, data: dict) -> dict:\n        """"""Split training data for current DP rank.""""""\n        total_lengths = [len(t) for t in data[""tokens""]]\n        data[""total_lengths""] = total_lengths\n\n        # Simple round-robin partitioning\n        partition = list(range(self.dp_rank, len(total_lengths), self.dp_size))\n\n        rollout_data = {""partition"": partition, ""total_lengths"": total_lengths}\n\n        for key in [\n            ""tokens"",\n            ""response_lengths"",\n            ""rewards"",\n            ""raw_reward"",\n            ""truncated"",\n            ""loss_masks"",\n            ""sample_indices"",\n        ]:\n            if key in data:\n                rollout_data[key] = [data[key][j] for j in partition]\n\n        return rollout_data\n\n    def _packed_data(self, rollout_data: dict) -> tuple[list[dict], list[int]]:\n        """"""Pack variable-length sequences for efficient processing.""""""\n        tokens = rollout_data[""tokens""]\n\n        packed_batches = []\n        mbs_size_list = []\n        local_batch_size = self.args.global_batch_size // self.dp_size\n\n        if self.args.use_dynamic_batch_size:\n            max_tokens = self.args.max_tokens_per_gpu\n            if self.cp_size > 1:\n                max_tokens = max_tokens * self.cp_size\n\n            for i in range(0, len(tokens), local_batch_size):\n                mbs_size_list.append(\n                    get_minimum_num_micro_batch_size(\n                        [len(t) for t in rollout_data[""tokens""][i : i + local_batch_size]],\n                        max_tokens,\n                    )\n                )\n            num_microbatches = torch.tensor(\n                mbs_size_list, dtype=torch.int, device=torch.cuda.current_device()\n            )\n            dist.all_reduce(num_microbatches, op=dist.ReduceOp.MAX, group=self.dp_group)\n            num_microbatches = num_microbatches.tolist()\n        else:\n            num_microbatches = [\n                self.args.global_batch_size // (self.args.micro_batch_size * self.dp_size)\n            ] * (len(tokens) // local_batch_size)\n\n        start = 0\n        for mbs_size in num_microbatches:\n            end = start + local_batch_size\n            # Create dummy advantages/returns for SFT (not used but required by pack_sequences)\n            dummy_advantages = [\n                torch.zeros(rollout_data[""response_lengths""][i])\n                for i in range(start, end)\n            ]\n            packed_batches.extend(\n                pack_sequences(\n                    rollout_data[""tokens""][start:end],\n                    rollout_data[""loss_masks""][start:end],\n                    rollout_data[""rewards""][start:end],\n                    rollout_data[""raw_reward""][start:end],\n                    rollout_data[""response_lengths""][start:end],\n                    dummy_advantages,  # advantages\n                    dummy_advantages,  # returns\n                    num_packs=mbs_size,\n                )\n            )\n            start = end\n\n        grad_accum = list(accumulate(num_microbatches))\n        return packed_batches, grad_accum\n\n    def _get_model_inputs_args(self, packed_sequence: dict) -> dict:\n        """"""Prepare model input arguments from packed sequence.""""""\n        input_ids = packed_sequence[""tokens""].unsqueeze(0)\n        position_ids = packed_sequence[""position_ids""].unsqueeze(0)\n\n        if self.cp_size > 1:\n            packed_sequence = pad_packed_sequence_with_cp(packed_sequence, self.cp_size)\n\n            if not packed_sequence[""cu_seqlens""].is_cuda:\n                packed_sequence[""cu_seqlens""] = packed_sequence[""cu_seqlens""].cuda()\n            cu_seqlens = packed_sequence[""cu_seqlens""]\n            update_ring_flash_attn_params(cu_seqlens, self.cp_group)\n\n            input_ids = torch.chunk(\n                packed_sequence[""tokens""].unsqueeze(0), self.cp_size, dim=1\n            )[self.cp_rank]\n            position_ids = torch.chunk(\n                packed_sequence[""position_ids""].unsqueeze(0), self.cp_size, dim=1\n            )[self.cp_rank]\n\n        model_args = {\n            ""input_ids"": input_ids,\n            ""position_ids"": position_ids,\n            ""attention_mask"": None,\n        }\n\n        if packed_sequence.get(""multimodal_inputs""):\n            model_args.update(packed_sequence[""multimodal_inputs""])\n\n        return model_args\n\n    def _compute_sft_loss(self, unpacked_batches: list[dict], logits: torch.Tensor):\n        """"""Compute SFT loss (negative log likelihood).""""""\n        loss_masks = [\n            batch[""loss_masks""].to(device=logits.device) for batch in unpacked_batches\n        ]\n        response_lengths = [batch[""response_lengths""] for batch in unpacked_batches]\n        log_probs = torch.cat(\n            [batch[""cur_log_probs""] for batch in unpacked_batches], dim=0\n        )\n        loss = -sum_of_sample_mean(log_probs, response_lengths, loss_masks)\n\n        if log_probs.numel() == 0:\n            loss += 0 * logits.sum()\n\n        return loss, {""loss"": loss.detach()}\n\n    def _train_step(\n        self,\n        packed_batch: dict,\n        reported_accum: dict,\n        mbs_id: int,\n        grad_accum: list[int],\n    ):\n        """"""Execute one training step.""""""\n        # Prepare model inputs\n        model_args = self._get_model_inputs_args(packed_batch)\n        logits = self.model(**model_args).logits.squeeze(0).float()\n\n        # Compute log probs and entropy (unified for both CP and non-CP modes)\n        log_probs, entropy_result = get_logprob_and_entropy_with_cp(\n            logits=logits,\n            target_tokens=packed_batch[""tokens""],\n            cp_rank=self.cp_rank,\n            cp_size=self.cp_size,\n            cp_group=self.cp_group,\n            model_input_ids=model_args[""input_ids""],\n            allow_compile=not self.args.true_on_policy_mode,\n            temperature=self.args.rollout_temperature,\n        )\n        packed_batch[""cur_log_probs""] = log_probs\n        packed_batch[""entropy""] = entropy_result\n\n        unpacked_batches = unpack_sequences(packed_batch)\n        loss, reported = self._compute_sft_loss(unpacked_batches, logits)\n\n        # Scale loss for gradient accumulation\n        loss = loss * self.dp_size / self.args.global_batch_size\n        loss.backward()\n\n        # Accumulate reported metrics (store tensors for later mean)\n        for k, v in reported.items():\n            reported_accum.setdefault(k, []).append(v)\n\n        if (mbs_id + 1) in grad_accum:\n            # TODO: check if the grad norm is global grad norm.\n            grad_norm = torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.args.clip_grad)\n            # the grad norm used to be of DTensor\n            grad_norm = float(grad_norm)\n\n            self.optimizer.step()\n            # Update learning rate\n            self.lr_scheduler.step()\n            self.optimizer.zero_grad(set_to_none=True)\n            # Aggregate logs\n            aggregated = {k: torch.stack(v).sum().item() for k, v in reported_accum.items()}\n            # TODO: change this, this is slow.\n            reduced_aggregated = [None] * self.dp_size\n            dist.all_gather_object(reduced_aggregated, aggregated, group=self.dp_group)\n            aggregated = {}\n            for k in reported_accum.keys():\n                aggregated[k] = sum([r[k] for r in reduced_aggregated]) / (self.args.global_batch_size)\n            reported_accum.clear()\n            if dist.get_rank() == 0:\n                log_dict = {\n                    f""train/{k}"": (val.item() if torch.is_tensor(val) else val) for k, val in aggregated.items()\n                }\n                log_dict[""train/grad_norm""] = grad_norm\n\n                # Log learning rate per parameter group; use scheduler's last computed LRs\n                lr_values = self.lr_scheduler.get_last_lr()\n                for gid, _group in enumerate(self.optimizer.param_groups):\n                    log_dict[f""train/lr_{gid}""] = lr_values[gid]\n\n                logger.info(f""step {self.global_step}: {log_dict}"")\n                log_dict[""train/step""] = self.global_step\n                tracking_utils.log(self.args, log_dict, step_key=""train/step"")\n            self.global_step += 1\n\n    def train_one_rollout(self, rollout_id: int):\n        """"""Execute one rollout's worth of training.""""""\n        self.model.train()\n        samples = self.generate_sft_rollout(rollout_id, self.data_source)\n\n        train_data = self._convert_samples_to_train_data(samples)\n\n        rollout_data = self._split_train_data_by_dp(train_data)\n\n        packed_batches, grad_accum = self._packed_data(rollout_data)\n\n        if len(grad_accum) == 0:\n            logger.warning(f""[Rank {dist.get_rank()}] No batches to train on rollout {rollout_id}"")\n            return\n\n        with timer(""actor_train""):\n            reported_accum = {}\n            self.optimizer.zero_grad(set_to_none=True)\n\n            for mbs_id, packed_batch in enumerate(\n                tqdm(packed_batches, desc=""actor_train"", disable=dist.get_rank() != 0)\n            ):\n                self._train_step(packed_batch, reported_accum, mbs_id, grad_accum)\n\n        self.prof.step(rollout_id=rollout_id)\n\n    def calculate_val_loss(self, rollout_id: int):\n        """"""Calculate validation loss over `args.val_steps`.""""""\n        self.model.eval()\n        reported_accum = {}\n        for v_step in tqdm(range(self.args.val_steps), desc=""actor_val"", disable=dist.get_rank() != 0):\n            samples = self.generate_sft_rollout(rollout_id, self.val_data_source)\n            val_data = self._convert_samples_to_train_data(samples)\n            rollout_data = self._split_train_data_by_dp(val_data)\n            packed_batches, accum = self._packed_data(rollout_data)\n\n            if len(accum) == 0:\n                logger.warning(f""[Rank {dist.get_rank()}] No batches to validate on rollout {rollout_id}, validation step {v_step}"")\n                return\n\n            for mbs_id, packed_batch in enumerate(packed_batches):\n                reported = self._val_step(packed_batch)\n                for k, v in reported.items():\n                    reported_accum.setdefault(k, []).append(v)\n\n        aggregated = {k: torch.stack(v).sum().item() for k, v in reported_accum.items()}\n        # TODO: change this, this is slow.\n        reduced_aggregated = [None] * self.dp_size\n        dist.all_gather_object(reduced_aggregated, aggregated, group=self.dp_group)\n        aggregated = {}\n        for k in reported_accum.keys():\n            aggregated[k] = sum([r[k] for r in reduced_aggregated]) / (self.args.global_batch_size * self.args.val_steps)\n        reported_accum.clear()\n        if dist.get_rank() == 0:\n            log_dict = {\n                f""val/{k}"": (val.item() if torch.is_tensor(val) else val) for k, val in aggregated.items()\n            }\n            logger.info(f""step {self.global_step}: {log_dict}"")\n            log_dict[""val/step""] = self.global_step\n            tracking_utils.log(self.args, log_dict, step_key=""val/step"")\n\n    def _val_step(self, packed_batch):\n        model_args = self._get_model_inputs_args(packed_batch)\n        with torch.no_grad():\n            logits = self.model(**model_args).logits.squeeze(0).float()\n\n        # Compute log probs and entropy (unified for both CP and non-CP modes)\n        log_probs, entropy_result = get_logprob_and_entropy_with_cp(\n            logits=logits,\n            target_tokens=packed_batch[""tokens""],\n            cp_rank=self.cp_rank,\n            cp_size=self.cp_size,\n            cp_group=self.cp_group,\n            model_input_ids=model_args[""input_ids""],\n            allow_compile=not self.args.true_on_policy_mode,\n            temperature=self.args.rollout_temperature,\n        )\n        packed_batch[""cur_log_probs""] = log_probs\n        packed_batch[""entropy""] = entropy_result\n\n        unpacked_batches = unpack_sequences(packed_batch)\n        _, reported = self._compute_sft_loss(unpacked_batches, logits)\n        return reported\n\n\n    def save_model(self, iteration: int):\n        """"""Save model checkpoint.""""""\n        if self.args.save is None:\n            return\n            \n        keys_filter = None\n        if self.args.use_lora:\n            keys_filter = lambda k: ""lora_"" in k\n            \n        checkpoint.save(self, iteration, keys_filter=keys_filter)\n        \n        if self.args.rollout_global_dataset:\n            self.data_source.save(iteration)\n\n    def train(self):\n        """"""Main training loop.""""""\n        logger.info(\n            f""[Rank {dist.get_rank()}] Starting training: ""\n            f""rollout_id {self.args.start_rollout_id} -> {self.args.num_rollout}""\n        )\n        if self.args.val_prompt_data:\n            assert self.args.val_interval > 0, f""val_interval must be greater than 0 when val_prompt_data is provided, got {self.args.val_interval}""\n            assert self.args.val_steps > 0, f""val_steps must be greater than 0 when val_prompt_data is provided, got {self.args.val_steps}""\n\n        # calculate val loss at the beginning of training\n        if self.args.val_prompt_data and self.args.start_rollout_id == 0:\n            self.calculate_val_loss(rollout_id=0)\n\n        for rollout_id in range(self.args.start_rollout_id, self.args.num_rollout):\n            self.train_one_rollout(rollout_id)\n\n            # Save checkpoint periodically\n            if should_run_periodic_action(\n                rollout_id, self.args.save_interval, self.num_rollout_per_epoch\n            ):\n                self.save_model(rollout_id)\n\n            # Calculate val loss periodically\n            if self.args.val_prompt_data and should_run_periodic_action(rollout_id, self.args.val_interval):\n                self.calculate_val_loss(rollout_id)\n\n        logger.info(f""[Rank {dist.get_rank()}] Training completed!"")\n\n\ndef set_sft_defaults(args: Namespace) -> Namespace:\n    """"""Set default values appropriate for SFT training.""""""\n    if not hasattr(args, ""loss_type"") or args.loss_type is None:\n        args.loss_type = ""sft_loss""\n\n    if not hasattr(args, ""advantage_estimator""):\n        args.advantage_estimator = None\n\n    args.offload_train = False\n    args.offload_rollout = False\n    args.colocate = False\n\n    return args\n\n\ndef main():\n    configure_logger()\n\n    args = parse_args()\n\n    args = set_sft_defaults(args)\n\n    trainer = SFTTrainer(args)\n    trainer.train()\n\n\nif __name__ == ""__main__"":\n    main()\n\n\n",python,tab
+5,55469,"scripts/run-sft-torchrun.sh",0,0,"#!/bin/bash\n#\n# Ray-free SFT Training Script\n#\nexport PYTHONUNBUFFERED=1\nexport PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True\n# FIXME(f.srambical): this is hardcoded for now\nGPUS_PER_NODE=${SLURM_GPUS_ON_NODE}\nNUM_NODES=${SLURM_JOB_NUM_NODES}\nNODE_RANK=${SLURM_NODEID}\nMASTER_ADDR=${MASTER_ADDR:-$(scontrol show hostnames $SLURM_JOB_NODELIST | head -n 1)}\n\nNVLINK_COUNT=$(nvidia-smi | grep -o ""NVLink"" | wc -l)\nif [ ""$NVLINK_COUNT"" -gt 0 ]; then\n    HAS_NVLINK=1\nelse\n    HAS_NVLINK=0\nfi\necho ""HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)""\n\nexport NCCL_DEBUG=INFO\nexport TORCH_DISTRIBUTED_DEBUG=INFO\n\nSCRIPT_DIR=""$(cd -- ""$(dirname -- ""${BASH_SOURCE[0]}"")"" &>/dev/null && pwd)""\n\nRUN_ID=${RUN_ID:-""run_$(date +%Y%m%d_%H%M%S)""}\nLOAD_PATH=""/fast/project/HFMI_SynergyUnit/tab_model/huggingface/Qwen3-0.6B""\nSAVE_PATH=""/fast/project/HFMI_SynergyUnit/tab_model/huggingface/shared_data/${RUN_ID}/checkpoints""\n\nCKPT_ARGS=(\n   --hf-checkpoint /fast/project/HFMI_SynergyUnit/tab_model/huggingface/Qwen3-0.6B\n   --load ${LOAD_PATH}\n   --ref-load /fast/project/HFMI_SynergyUnit/tab_model/huggingface/Qwen3-0.6B\n   --save ${SAVE_PATH}\n   --save-interval 1000\n)\n\nSFT_ARGS=(\n   --rollout-function-path miles.rollout.sft_rollout.generate_rollout\n   --prompt-data /fast/project/HFMI_SynergyUnit/tab_model/huggingface/nemo_hf_part_jsonl_4k_tokens.jsonl\n   --val-prompt-data /fast/project/HFMI_SynergyUnit/tab_model/huggingface/nemo_hf_part_jsonl_4k_tokens_validation.jsonl\n   --val-interval 1000\n   --val-steps 100\n   --input-key messages\n   --apply-chat-template\n   --rollout-shuffle\n   --num-rollout 10000\n   --rollout-batch-size 16\n   --global-batch-size 16\n\n   --loss-type sft_loss\n   --calculate-per-token-loss\n   --disable-compute-advantages-and-returns\n)\n\nLORA_ARGS=(\n    --use-lora\n    --lora-rank 8\n    --lora-alpha 16\n    --lora-dropout 0.0\n    --lora-target-modules q_proj v_proj\n)\n\nOPTIMIZER_ARGS=(\n   --optimizer adam\n   --lr 1e-4\n   --lr-decay-style WSD\n   --lr-wsd-decay-style linear\n   --lr-warmup-iters 500\n   --lr-decay-iters 10000\n   --lr-wsd-decay-iters 2000\n   --weight-decay 0.1\n   --adam-beta1 0.9\n   --adam-beta2 0.98\n)\n\nWANDB_ARGS=(\n   --use-wandb\n   --wandb-project crowd-pilot-miles\n   --wandb-team instant-uv\n   --wandb-group qwen3-0.6b-sft-torchrun\n)\n\nTRAIN_BACKEND_ARGS=(\n   --train-backend fsdp\n   --update-weight-buffer-size 536870912\n   --attn-implementation flash_attention_3\n)\n\nPERF_ARGS=(\n   --use-dynamic-batch-size\n   --max-tokens-per-gpu 9216\n)\n\nMISC_ARGS=(\n   --rollout-max-context-len 8192\n   --rollout-max-prompt-len 8000\n   --rollout-max-response-len 8192\n   --dump-details /fast/project/HFMI_SynergyUnit/tab_model/huggingface/shared_data/qwen3-600M-fsdp-1116-noref/dump_details\n)\n\ntorchrun \\n    --nproc_per_node=${GPUS_PER_NODE} \\n    --nnodes=${NUM_NODES} \\n    --node_rank=${NODE_RANK} \\n    --master_addr=${MASTER_ADDR} \\n    --master_port=${MASTER_PORT:-29500} \\n    train_sft.py \\n    ${CKPT_ARGS[@]} \\n    ${SFT_ARGS[@]} \\n    ${LORA_ARGS[@]} \\n    ${OPTIMIZER_ARGS[@]} \\n    ${WANDB_ARGS[@]} \\n    ${TRAIN_BACKEND_ARGS[@]} \\n    ${PERF_ARGS[@]} \\n    ${MISC_ARGS[@]}\n",shellscript,tab
+6,62658,"/home/franz.srambical/slurm/jobs/mihir/berlin/crowd-pilot/run_sqwen3-0.6b-sft-torch-crowd-code.sh",0,0,"#!/usr/bin/env bash\n\n#SBATCH --nodes=1\n#SBATCH --ntasks-per-node=1\n#SBATCH --time=24:00:00\n#SBATCH --cpus-per-task=5\n#SBATCH --gres=gpu:2\n#SBATCH --mem=600GB\n#SBATCH --signal=USR1@60\n#SBATCH --output=/fast/project/HFMI_SynergyUnit/mihir/logs/train/qwen3-0.6b-sft-torch-crowd-code/%x_%j.log\n#SBATCH --error=/fast/project/HFMI_SynergyUnit/mihir/logs/train/qwen3-0.6b-sft-torch-crowd-code/%x_%j.log\n#SBATCH --job-name=run_qwen3-0.6b-sft-torch-crowd-code\n\n\ntrap 'echo ""Job killed by signal""' USR1\n\ncat $0\nmodule load CUDA/12.8\n\nsource .venv/bin/activate\n\nsrun sh scripts/run-qwen3-0.6B-torch-sft-crowd-code.sh",shellscript,tab
+7,63969,"scripts/run-sft-torchrun.sh",0,0,"",shellscript,tab
+8,79858,"train_sft.py",0,0,"",python,tab
+9,124769,"train_sft.py",11340,0,"",python,selection_command
+10,130824,"train_sft.py",11326,0,"",python,selection_command
+11,130929,"train_sft.py",11325,0,"",python,selection_command
+12,131497,"train_sft.py",11326,0,"",python,selection_command
+13,132309,"train_sft.py",11325,0,"",python,selection_command
+14,133776,"train_sft.py",11326,0,"",python,selection_command
+15,135869,"train_sft.py",11325,0,"",python,selection_command
+16,135989,"train_sft.py",11320,0,"",python,selection_command
+17,136409,"train_sft.py",11038,0,"",python,selection_command
+18,136789,"train_sft.py",11030,0,"",python,selection_command
+19,137009,"train_sft.py",10884,0,"",python,selection_command
+20,137689,"train_sft.py",10830,0,"",python,selection_command
+21,138325,"train_sft.py",10836,0,"",python,selection_command
+22,138437,"train_sft.py",10838,0,"",python,selection_command
+23,138678,"train_sft.py",10842,0,"",python,selection_command
+24,139969,"train_sft.py",10843,0,"",python,selection_command
+25,140499,"train_sft.py",8279,0,"",python,selection_command
+26,143937,"train_sft.py",10843,0,"",python,selection_command
+27,145273,"train_sft.py",10889,0,"",python,selection_command
+28,145529,"train_sft.py",10912,0,"",python,selection_command
+29,145593,"train_sft.py",10936,0,"",python,selection_command
+30,145593,"train_sft.py",10962,0,"",python,selection_command
+31,145594,"train_sft.py",11014,0,"",python,selection_command
+32,145649,"train_sft.py",11016,0,"",python,selection_command
+33,145688,"train_sft.py",11038,0,"",python,selection_command
+34,145745,"train_sft.py",11044,0,"",python,selection_command
+35,145745,"train_sft.py",11066,0,"",python,selection_command
+36,145809,"train_sft.py",11111,0,"",python,selection_command
+37,145816,"train_sft.py",11230,0,"",python,selection_command
+38,146297,"train_sft.py",11324,0,"",python,selection_command
+39,146617,"train_sft.py",11325,0,"",python,selection_command
+40,147097,"train_sft.py",11326,0,"",python,selection_command
+41,147537,"train_sft.py",11326,1,"g",python,selection_command
+42,147709,"train_sft.py",11326,29,"gradient_checkpointing_enable",python,selection_command
+43,148129,"train_sft.py",11354,0,"",python,selection_command
+44,158417,"train_sft.py",11326,0,"",python,selection_command
+45,158648,"train_sft.py",11325,0,"",python,selection_command
+46,159097,"train_sft.py",11326,0,"",python,selection_command
+47,159281,"train_sft.py",11325,0,"",python,selection_command
+48,159368,"train_sft.py",11325,1,".",python,selection_command
+49,159769,"train_sft.py",11325,0,"",python,selection_command
+50,160204,"train_sft.py",11326,0,"",python,selection_command
+51,160205,"train_sft.py",11326,1,"g",python,selection_command
+52,160337,"train_sft.py",11326,29,"gradient_checkpointing_enable",python,selection_command
+53,160496,"train_sft.py",11326,31,"gradient_checkpointing_enable()",python,selection_command
+54,161289,"train_sft.py",11356,0,"",python,selection_command
+55,1742088,"train_sft.py",11303,54,"            self.model.gradient_checkpointing_enable()",python,selection_command
+56,1831498,"train_sft.py",11356,0,"",python,selection_command
+57,2470031,"train_sft.py",0,29058,"#!/usr/bin/env python3\n""""""\nRay-free SFT Training Script for Miles.\n\nThis script provides a simplified training path for Supervised Fine-Tuning (SFT)\nthat bypasses Ray entirely and uses torchrun for distributed training.\n\nUsage:\n    torchrun --nproc_per_node=2 train_sft.py \\n        --hf-checkpoint /path/to/model \\n        --prompt-data /path/to/data.parquet \\n        --input-key messages \\n        --apply-chat-template \\n        ...\n\nThis is equivalent to the Ray-based SFT with --debug-train-only, but without\nthe Ray overhead.\n""""""\n\nimport logging\nimport os\nfrom argparse import Namespace\nfrom datetime import timedelta\nfrom itertools import accumulate\n\nimport torch\nimport torch.distributed as dist\nfrom torch.distributed.device_mesh import init_device_mesh\nfrom tqdm import tqdm\nfrom transformers import AutoConfig\n\nfrom ring_flash_attn import substitute_hf_flash_attn, update_ring_flash_attn_params\n\nfrom miles.models.peft import LoRAConfig, apply_lora\nfrom miles.backends.fsdp_utils import checkpoint\nfrom miles.backends.fsdp_utils.actor import (\n    apply_fsdp2,\n    get_logprob_and_entropy_with_cp,\n    sum_of_sample_mean,\n)\nfrom miles.backends.fsdp_utils.data_packing import (\n    pack_sequences,\n    pad_packed_sequence_with_cp,\n    unpack_sequences,\n)\nfrom miles.backends.fsdp_utils.lr_scheduler import get_lr_scheduler\nfrom miles.rollout.data_source import RolloutDataSource\nfrom miles.utils import tracking_utils\nfrom miles.utils.arguments import parse_args\nfrom miles.utils.data import get_minimum_num_micro_batch_size\nfrom miles.utils.distributed_utils import get_gloo_group, init_gloo_group\nfrom miles.utils.logging_utils import configure_logger\nfrom miles.utils.mask_utils import MultiTurnLossMaskGenerator\nfrom miles.utils.misc import should_run_periodic_action\nfrom miles.utils.processing_utils import load_processor, load_tokenizer\nfrom miles.utils.profile_utils import TrainProfiler\nfrom miles.utils.timer import timer\nfrom miles.utils.tracking_utils import init_tracking\nfrom miles.utils.types import Sample\n\nlogger = logging.getLogger(__name__)\n\n\nclass SFTTrainer:\n    """"""\n    A simplified trainer for SFT that runs without Ray.\n\n    This class combines the functionality of:\n    - FSDPTrainRayActor (model initialization, FSDP wrapping, training)\n    - RolloutManager (data loading via generate_rollout)\n    - The main training loop from train.py\n    """"""\n\n    def __init__(self, args: Namespace):\n        self.args = args\n        self.device = torch.device(""cuda"")\n\n        self._init_distributed()\n\n        self._setup_device_mesh()\n\n        torch.manual_seed(args.seed)\n\n        self._enable_true_on_policy_optimizations()\n\n        if dist.get_rank() == 0:\n            init_tracking(args, primary=True)\n\n        self._load_tokenizer_and_config()\n\n        self._init_data_source()\n\n        self._init_model()\n\n        self._init_optimizer()\n\n        self._load_checkpoint()\n\n        self.prof = TrainProfiler(args)\n        self.prof.on_init_end()\n\n        logger.info(f""[Rank {dist.get_rank()}] SFTTrainer initialized successfully"")\n\n    def _init_distributed(self):\n        """"""Initialize distributed training.""""""\n        # torchrun sets these environment variables\n        local_rank = int(os.environ.get(""LOCAL_RANK"", 0))\n        torch.cuda.set_device(f""cuda:{local_rank}"")\n\n        backend = self.args.distributed_backend\n        dist.init_process_group(\n            backend=backend,\n            timeout=timedelta(minutes=self.args.distributed_timeout_minutes),\n        )\n        init_gloo_group()\n\n        self.args.rank = dist.get_rank()\n        self.args.world_size = dist.get_world_size()\n\n        logger.info(\n            f""[Rank {self.args.rank}] Distributed initialized: ""\n            f""world_size={self.args.world_size}, local_rank={local_rank}""\n        )\n\n    def _setup_device_mesh(self):\n        """"""Setup device mesh for FSDP (no context parallelism for SFT).""""""\n        world_size = dist.get_world_size()\n        rank = dist.get_rank()\n\n        self.cp_size = self.args.context_parallel_size\n        self.dp_size = world_size // self.cp_size\n\n        self.mesh = init_device_mesh(\n            ""cuda"",\n            mesh_shape=(self.dp_size, self.cp_size),\n            mesh_dim_names=(""dp"", ""cp""),\n        )\n\n        self.dp_group = self.mesh.get_group(""dp"")\n        self.cp_group = self.mesh.get_group(""cp"")\n        self.dp_mesh = self.mesh[""dp""]\n\n        self.dp_rank = rank // self.cp_size\n        self.cp_rank = rank % self.cp_size\n\n        logger.info(\n            f""[Rank {rank}] Device mesh: dp_size={self.dp_size}, cp_size={self.cp_size}, ""\n            f""dp_rank={self.dp_rank}, cp_rank={self.cp_rank}""\n        )\n\n        # Setup Ring Flash Attention with CP group from mesh (only when cp_size > 1)\n        if self.cp_size > 1:\n            substitute_hf_flash_attn(self.cp_group, heads_k_stride=1)\n            logger.info(f""[Rank {rank}] CP initialized via device mesh"")\n\n    def _enable_true_on_policy_optimizations(self):\n        """"""Enable true on-policy optimizations or apply MoE patches.""""""\n        if self.args.true_on_policy_mode:\n            from sglang.srt.batch_invariant_ops import enable_batch_invariant_mode\n\n            from miles.backends.fsdp_utils.models.qwen3_moe import (\n                apply_true_on_policy_patch_for_qwen3_moe,\n            )\n\n            logger.info(""SFTTrainer: enabling batch_invariant_mode for true-on-policy"")\n            enable_batch_invariant_mode(\n                # In Qwen3, rope uses bmm; disabling makes it aligned\n                enable_bmm=False,\n            )\n\n            apply_true_on_policy_patch_for_qwen3_moe()\n        else:\n            from miles.backends.fsdp_utils.models.qwen3_moe_hf import apply_fsdp_moe_patch\n\n            apply_fsdp_moe_patch()\n\n    def _load_tokenizer_and_config(self):\n        """"""Load tokenizer and model config sequentially to avoid race conditions.""""""\n        for i in range(dist.get_world_size()):\n            if i == dist.get_rank():\n                self.hf_config = AutoConfig.from_pretrained(\n                    self.args.hf_checkpoint, trust_remote_code=True\n                )\n                self.tokenizer = load_tokenizer(\n                    self.args.hf_checkpoint, trust_remote_code=True\n                )\n                self.processor = None\n                if self.args.multimodal_keys:\n                    self.processor = load_processor(\n                        self.args.hf_checkpoint, trust_remote_code=True\n                    )\n            dist.barrier(group=get_gloo_group())\n\n        # Initialize loss mask generator for SFT\n        self.mask_generator = MultiTurnLossMaskGenerator(\n            self.tokenizer,\n            tokenizer_type=getattr(self.args, ""loss_mask_type"", None),\n        )\n\n    def _init_data_source(self):\n        """"""Initialize the data source for SFT training.""""""\n        self.data_source = RolloutDataSource(self.args, self.args.prompt_data)\n        self.val_data_source = None\n        if self.args.val_prompt_data is not None:\n            self.val_data_source = RolloutDataSource(self.args, self.args.val_prompt_data)\n\n        # Calculate num_rollout from dataset size\n        if self.args.num_rollout is None:\n            num_rollout_per_epoch = len(self.data_source.dataset) // self.args.rollout_batch_size\n            self.args.num_rollout = num_rollout_per_epoch * self.args.num_epoch\n            self.num_rollout_per_epoch = num_rollout_per_epoch\n        else:\n            self.num_rollout_per_epoch = None\n\n        if getattr(self.args, ""start_rollout_id"", None) is None:\n            self.args.start_rollout_id = 0\n\n        logger.info(\n            f""[Rank {dist.get_rank()}] Data source initialized: ""\n            f""dataset_size={len(self.data_source.dataset)}, ""\n            f""num_rollout={self.args.num_rollout}""\n        )\n\n    def _get_init_weight_context_manager(self):\n        """"""Get context manager for model initialization.""""""\n        from accelerate import init_empty_weights\n\n        use_meta_tensor = not self.hf_config.tie_word_embeddings\n\n        def cpu_init_weights():\n            return torch.device(""cpu"")\n\n        if use_meta_tensor:\n            return init_empty_weights if dist.get_rank() != 0 else cpu_init_weights\n        else:\n            return cpu_init_weights\n\n    def _fsdp2_load_full_state_dict(self, model, full_state, device_mesh, cpu_offload):\n        """"""Load full state dict into FSDP2 model with broadcast from rank 0.""""""\n        from torch.distributed.checkpoint.state_dict import (\n            StateDictOptions,\n            set_model_state_dict,\n        )\n\n        if dist.get_rank() == 0:\n            model = model.to(device=torch.cuda.current_device(), non_blocking=True)\n        else:\n            model = model.to_empty(device=torch.cuda.current_device())\n\n        is_cpu_offload = cpu_offload is not None\n        options = StateDictOptions(\n            full_state_dict=True, cpu_offload=is_cpu_offload, broadcast_from_rank0=True\n        )\n\n        set_model_state_dict(model, full_state, options=options)\n\n        for _name, buf in model.named_buffers():\n            dist.broadcast(buf, src=0)\n\n        if is_cpu_offload:\n            model.to(""cpu"", non_blocking=True)\n            for buf in model.buffers():\n                buf.data = buf.data.to(torch.cuda.current_device())\n\n        return model\n\n    def _get_model_cls(self):\n        """"""Get the appropriate model class based on config.""""""\n        if hasattr(self.hf_config, ""vision_config""):\n            from transformers import AutoModelForImageTextToText\n\n            return AutoModelForImageTextToText\n        else:\n            from transformers import AutoModelForCausalLM\n\n            return AutoModelForCausalLM\n\n    def _init_model(self):\n        """"""Initialize and wrap model with FSDP.""""""\n        self.fsdp_cpu_offload = getattr(self.args, ""fsdp_cpu_offload"", False)\n\n        init_context = self._get_init_weight_context_manager()\n\n        with init_context():\n            model = self._get_model_cls().from_pretrained(\n                self.args.hf_checkpoint,\n                trust_remote_code=True,\n                attn_implementation=self.args.attn_implementation,\n            )\n\n            if self.args.use_lora:\n                lora_config = LoRAConfig(\n                    lora_rank=self.args.lora_rank,\n                    lora_alpha=self.args.lora_alpha,\n                    lora_dropout=self.args.lora_dropout,\n                    target_modules=self.args.lora_target_modules,\n                )\n                model = apply_lora(model, lora_config)\n                logger.info(f""[Rank {dist.get_rank()}] Applied LoRA: {lora_config}"")\n\n        model.train()\n        full_state = model.state_dict()\n\n        model = apply_fsdp2(\n            model, mesh=self.dp_mesh, cpu_offload=self.fsdp_cpu_offload, args=self.args\n        )\n\n        model = self._fsdp2_load_full_state_dict(\n            model,\n            full_state,\n            self.dp_mesh,\n            cpu_offload=True if self.fsdp_cpu_offload else None,\n        )\n\n        self.model = model\n\n        if self.args.gradient_checkpointing:\n            # FIXME: Conceptually, gradient checkpointing should be compatible with LoRA, but we don't support it yet.\n            assert not self.args.use_lora, ""Gradient checkpointing is incompatible with LoRA""\n            self.model.gradient_checkpointing_enable()\n\n        logger.info(f""[Rank {dist.get_rank()}] Model initialized with FSDP"")\n\n    def _init_optimizer(self):\n        """"""Initialize optimizer and learning rate scheduler.""""""\n        trainable_params = [p for p in self.model.parameters() if p.requires_grad]\n        \n        if self.args.use_lora:\n            total_params = sum(p.numel() for p in self.model.parameters())\n            trainable_count = sum(p.numel() for p in trainable_params)\n            logger.info(\n                f""[Rank {dist.get_rank()}] LoRA: {trainable_count:,} trainable params ""\n                f""out of {total_params:,} total ({100 * trainable_count / total_params:.2f}%)""\n            )\n        \n        if self.args.optimizer == ""adam"":\n            self.optimizer = torch.optim.AdamW(\n                trainable_params,\n                lr=self.args.lr,\n                betas=(self.args.adam_beta1, self.args.adam_beta2),\n                eps=self.args.adam_eps,\n                weight_decay=self.args.weight_decay,\n            )\n        else:\n            raise ValueError(f""Unsupported optimizer: {self.args.optimizer}"")\n\n        self.lr_scheduler = get_lr_scheduler(self.args, self.optimizer)\n        self.global_step = 0\n        self.micro_step = 0\n\n    def _load_checkpoint(self):\n        """"""Load checkpoint if available.""""""\n        checkpoint_payload = checkpoint.load(self)\n        checkpoint.finalize_load(self, checkpoint_payload)\n        \n        if self.args.rollout_global_dataset and self.args.start_rollout_id > 0:\n            self.data_source.load(self.args.start_rollout_id - 1)\n\n    def generate_sft_rollout(self, rollout_id: int, data_source: RolloutDataSource) -> list[Sample]:\n        """"""Generate SFT rollout data (tokenize and create loss masks).""""""\n        samples = data_source.get_samples(self.args.rollout_batch_size)\n\n        result = []\n        for i, (sample,) in enumerate(samples):\n            messages = sample.prompt\n            token_ids, loss_mask = self.mask_generator.get_loss_mask(messages)\n            response_length = self.mask_generator.get_response_lengths([loss_mask])[0]\n\n            sample.tokens = token_ids\n            sample.response_length = response_length\n            sample.reward = 0\n            sample.loss_mask = loss_mask[-response_length:]\n            result.append(sample)\n\n            if i == 0 and rollout_id == 0 and dist.get_rank() == 0:\n                logger.info(\n                    f""SFT rollout sample: tokens_len={len(token_ids)}, ""\n                    f""response_length={response_length}""\n                )\n\n        return result\n\n    def _convert_samples_to_train_data(self, samples: list[Sample]) -> dict:\n        """"""Convert samples to training data format.""""""\n        train_data = {\n            ""tokens"": [sample.tokens for sample in samples],\n            ""response_lengths"": [sample.response_length for sample in samples],\n            ""rewards"": [0.0 for _ in samples],\n            ""raw_reward"": [0.0 for _ in samples],\n            ""truncated"": [0 for _ in samples],\n            ""sample_indices"": [sample.index for sample in samples],\n        }\n\n        loss_masks = []\n        for sample in samples:\n            if sample.loss_mask is None:\n                sample.loss_mask = [1] * sample.response_length\n            loss_masks.append(sample.loss_mask)\n        train_data[""loss_masks""] = loss_masks\n\n        return train_data\n\n    def _split_train_data_by_dp(self, data: dict) -> dict:\n        """"""Split training data for current DP rank.""""""\n        total_lengths = [len(t) for t in data[""tokens""]]\n        data[""total_lengths""] = total_lengths\n\n        # Simple round-robin partitioning\n        partition = list(range(self.dp_rank, len(total_lengths), self.dp_size))\n\n        rollout_data = {""partition"": partition, ""total_lengths"": total_lengths}\n\n        for key in [\n            ""tokens"",\n            ""response_lengths"",\n            ""rewards"",\n            ""raw_reward"",\n            ""truncated"",\n            ""loss_masks"",\n            ""sample_indices"",\n        ]:\n            if key in data:\n                rollout_data[key] = [data[key][j] for j in partition]\n\n        return rollout_data\n\n    def _packed_data(self, rollout_data: dict) -> tuple[list[dict], list[int]]:\n        """"""Pack variable-length sequences for efficient processing.""""""\n        tokens = rollout_data[""tokens""]\n\n        packed_batches = []\n        mbs_size_list = []\n        local_batch_size = self.args.global_batch_size // self.dp_size\n\n        if self.args.use_dynamic_batch_size:\n            max_tokens = self.args.max_tokens_per_gpu\n            if self.cp_size > 1:\n                max_tokens = max_tokens * self.cp_size\n\n            for i in range(0, len(tokens), local_batch_size):\n                mbs_size_list.append(\n                    get_minimum_num_micro_batch_size(\n                        [len(t) for t in rollout_data[""tokens""][i : i + local_batch_size]],\n                        max_tokens,\n                    )\n                )\n            num_microbatches = torch.tensor(\n                mbs_size_list, dtype=torch.int, device=torch.cuda.current_device()\n            )\n            dist.all_reduce(num_microbatches, op=dist.ReduceOp.MAX, group=self.dp_group)\n            num_microbatches = num_microbatches.tolist()\n        else:\n            num_microbatches = [\n                self.args.global_batch_size // (self.args.micro_batch_size * self.dp_size)\n            ] * (len(tokens) // local_batch_size)\n\n        start = 0\n        for mbs_size in num_microbatches:\n            end = start + local_batch_size\n            # Create dummy advantages/returns for SFT (not used but required by pack_sequences)\n            dummy_advantages = [\n                torch.zeros(rollout_data[""response_lengths""][i])\n                for i in range(start, end)\n            ]\n            packed_batches.extend(\n                pack_sequences(\n                    rollout_data[""tokens""][start:end],\n                    rollout_data[""loss_masks""][start:end],\n                    rollout_data[""rewards""][start:end],\n                    rollout_data[""raw_reward""][start:end],\n                    rollout_data[""response_lengths""][start:end],\n                    dummy_advantages,  # advantages\n                    dummy_advantages,  # returns\n                    num_packs=mbs_size,\n                )\n            )\n            start = end\n\n        grad_accum = list(accumulate(num_microbatches))\n        return packed_batches, grad_accum\n\n    def _get_model_inputs_args(self, packed_sequence: dict) -> dict:\n        """"""Prepare model input arguments from packed sequence.""""""\n        input_ids = packed_sequence[""tokens""].unsqueeze(0)\n        position_ids = packed_sequence[""position_ids""].unsqueeze(0)\n\n        if self.cp_size > 1:\n            packed_sequence = pad_packed_sequence_with_cp(packed_sequence, self.cp_size)\n\n            if not packed_sequence[""cu_seqlens""].is_cuda:\n                packed_sequence[""cu_seqlens""] = packed_sequence[""cu_seqlens""].cuda()\n            cu_seqlens = packed_sequence[""cu_seqlens""]\n            update_ring_flash_attn_params(cu_seqlens, self.cp_group)\n\n            input_ids = torch.chunk(\n                packed_sequence[""tokens""].unsqueeze(0), self.cp_size, dim=1\n            )[self.cp_rank]\n            position_ids = torch.chunk(\n                packed_sequence[""position_ids""].unsqueeze(0), self.cp_size, dim=1\n            )[self.cp_rank]\n\n        model_args = {\n            ""input_ids"": input_ids,\n            ""position_ids"": position_ids,\n            ""attention_mask"": None,\n        }\n\n        if packed_sequence.get(""multimodal_inputs""):\n            model_args.update(packed_sequence[""multimodal_inputs""])\n\n        return model_args\n\n    def _compute_sft_loss(self, unpacked_batches: list[dict], logits: torch.Tensor):\n        """"""Compute SFT loss (negative log likelihood).""""""\n        loss_masks = [\n            batch[""loss_masks""].to(device=logits.device) for batch in unpacked_batches\n        ]\n        response_lengths = [batch[""response_lengths""] for batch in unpacked_batches]\n        log_probs = torch.cat(\n            [batch[""cur_log_probs""] for batch in unpacked_batches], dim=0\n        )\n        loss = -sum_of_sample_mean(log_probs, response_lengths, loss_masks)\n\n        if log_probs.numel() == 0:\n            loss += 0 * logits.sum()\n\n        return loss, {""loss"": loss.detach()}\n\n    def _train_step(\n        self,\n        packed_batch: dict,\n        reported_accum: dict,\n        mbs_id: int,\n        grad_accum: list[int],\n    ):\n        """"""Execute one training step.""""""\n        # Prepare model inputs\n        model_args = self._get_model_inputs_args(packed_batch)\n        logits = self.model(**model_args).logits.squeeze(0).float()\n\n        # Compute log probs and entropy (unified for both CP and non-CP modes)\n        log_probs, entropy_result = get_logprob_and_entropy_with_cp(\n            logits=logits,\n            target_tokens=packed_batch[""tokens""],\n            cp_rank=self.cp_rank,\n            cp_size=self.cp_size,\n            cp_group=self.cp_group,\n            model_input_ids=model_args[""input_ids""],\n            allow_compile=not self.args.true_on_policy_mode,\n            temperature=self.args.rollout_temperature,\n        )\n        packed_batch[""cur_log_probs""] = log_probs\n        packed_batch[""entropy""] = entropy_result\n\n        unpacked_batches = unpack_sequences(packed_batch)\n        loss, reported = self._compute_sft_loss(unpacked_batches, logits)\n\n        # Scale loss for gradient accumulation\n        loss = loss * self.dp_size / self.args.global_batch_size\n        loss.backward()\n\n        # Accumulate reported metrics (store tensors for later mean)\n        for k, v in reported.items():\n            reported_accum.setdefault(k, []).append(v)\n\n        if (mbs_id + 1) in grad_accum:\n            # TODO: check if the grad norm is global grad norm.\n            grad_norm = torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.args.clip_grad)\n            # the grad norm used to be of DTensor\n            grad_norm = float(grad_norm)\n\n            self.optimizer.step()\n            # Update learning rate\n            self.lr_scheduler.step()\n            self.optimizer.zero_grad(set_to_none=True)\n            # Aggregate logs\n            aggregated = {k: torch.stack(v).sum().item() for k, v in reported_accum.items()}\n            # TODO: change this, this is slow.\n            reduced_aggregated = [None] * self.dp_size\n            dist.all_gather_object(reduced_aggregated, aggregated, group=self.dp_group)\n            aggregated = {}\n            for k in reported_accum.keys():\n                aggregated[k] = sum([r[k] for r in reduced_aggregated]) / (self.args.global_batch_size)\n            reported_accum.clear()\n            if dist.get_rank() == 0:\n                log_dict = {\n                    f""train/{k}"": (val.item() if torch.is_tensor(val) else val) for k, val in aggregated.items()\n                }\n                log_dict[""train/grad_norm""] = grad_norm\n\n                # Log learning rate per parameter group; use scheduler's last computed LRs\n                lr_values = self.lr_scheduler.get_last_lr()\n                for gid, _group in enumerate(self.optimizer.param_groups):\n                    log_dict[f""train/lr_{gid}""] = lr_values[gid]\n\n                logger.info(f""step {self.global_step}: {log_dict}"")\n                log_dict[""train/step""] = self.global_step\n                tracking_utils.log(self.args, log_dict, step_key=""train/step"")\n            self.global_step += 1\n\n    def train_one_rollout(self, rollout_id: int):\n        """"""Execute one rollout's worth of training.""""""\n        self.model.train()\n        samples = self.generate_sft_rollout(rollout_id, self.data_source)\n\n        train_data = self._convert_samples_to_train_data(samples)\n\n        rollout_data = self._split_train_data_by_dp(train_data)\n\n        packed_batches, grad_accum = self._packed_data(rollout_data)\n\n        if len(grad_accum) == 0:\n            logger.warning(f""[Rank {dist.get_rank()}] No batches to train on rollout {rollout_id}"")\n            return\n\n        with timer(""actor_train""):\n            reported_accum = {}\n            self.optimizer.zero_grad(set_to_none=True)\n\n            for mbs_id, packed_batch in enumerate(\n                tqdm(packed_batches, desc=""actor_train"", disable=dist.get_rank() != 0)\n            ):\n                self._train_step(packed_batch, reported_accum, mbs_id, grad_accum)\n\n        self.prof.step(rollout_id=rollout_id)\n\n    def calculate_val_loss(self, rollout_id: int):\n        """"""Calculate validation loss over `args.val_steps`.""""""\n        self.model.eval()\n        reported_accum = {}\n        for v_step in tqdm(range(self.args.val_steps), desc=""actor_val"", disable=dist.get_rank() != 0):\n            samples = self.generate_sft_rollout(rollout_id, self.val_data_source)\n            val_data = self._convert_samples_to_train_data(samples)\n            rollout_data = self._split_train_data_by_dp(val_data)\n            packed_batches, accum = self._packed_data(rollout_data)\n\n            if len(accum) == 0:\n                logger.warning(f""[Rank {dist.get_rank()}] No batches to validate on rollout {rollout_id}, validation step {v_step}"")\n                return\n\n            for mbs_id, packed_batch in enumerate(packed_batches):\n                reported = self._val_step(packed_batch)\n                for k, v in reported.items():\n                    reported_accum.setdefault(k, []).append(v)\n\n        aggregated = {k: torch.stack(v).sum().item() for k, v in reported_accum.items()}\n        # TODO: change this, this is slow.\n        reduced_aggregated = [None] * self.dp_size\n        dist.all_gather_object(reduced_aggregated, aggregated, group=self.dp_group)\n        aggregated = {}\n        for k in reported_accum.keys():\n            aggregated[k] = sum([r[k] for r in reduced_aggregated]) / (self.args.global_batch_size * self.args.val_steps)\n        reported_accum.clear()\n        if dist.get_rank() == 0:\n            log_dict = {\n                f""val/{k}"": (val.item() if torch.is_tensor(val) else val) for k, val in aggregated.items()\n            }\n            logger.info(f""step {self.global_step}: {log_dict}"")\n            log_dict[""val/step""] = self.global_step\n            tracking_utils.log(self.args, log_dict, step_key=""val/step"")\n\n    def _val_step(self, packed_batch):\n        model_args = self._get_model_inputs_args(packed_batch)\n        with torch.no_grad():\n            logits = self.model(**model_args).logits.squeeze(0).float()\n\n        # Compute log probs and entropy (unified for both CP and non-CP modes)\n        log_probs, entropy_result = get_logprob_and_entropy_with_cp(\n            logits=logits,\n            target_tokens=packed_batch[""tokens""],\n            cp_rank=self.cp_rank,\n            cp_size=self.cp_size,\n            cp_group=self.cp_group,\n            model_input_ids=model_args[""input_ids""],\n            allow_compile=not self.args.true_on_policy_mode,\n            temperature=self.args.rollout_temperature,\n        )\n        packed_batch[""cur_log_probs""] = log_probs\n        packed_batch[""entropy""] = entropy_result\n\n        unpacked_batches = unpack_sequences(packed_batch)\n        _, reported = self._compute_sft_loss(unpacked_batches, logits)\n        return reported\n\n\n    def save_model(self, iteration: int):\n        """"""Save model checkpoint.""""""\n        if self.args.save is None:\n            return\n            \n        keys_filter = None\n        if self.args.use_lora:\n            keys_filter = lambda k: ""lora_"" in k\n            \n        checkpoint.save(self, iteration, keys_filter=keys_filter)\n        \n        if self.args.rollout_global_dataset:\n            self.data_source.save(iteration)\n\n    def train(self):\n        """"""Main training loop.""""""\n        logger.info(\n            f""[Rank {dist.get_rank()}] Starting training: ""\n            f""rollout_id {self.args.start_rollout_id} -> {self.args.num_rollout}""\n        )\n        if self.args.val_prompt_data:\n            assert self.args.val_interval > 0, f""val_interval must be greater than 0 when val_prompt_data is provided, got {self.args.val_interval}""\n            assert self.args.val_steps > 0, f""val_steps must be greater than 0 when val_prompt_data is provided, got {self.args.val_steps}""\n\n        # calculate val loss at the beginning of training\n        if self.args.val_prompt_data and self.args.start_rollout_id == 0:\n            self.calculate_val_loss(rollout_id=0)\n\n        for rollout_id in range(self.args.start_rollout_id, self.args.num_rollout):\n            self.train_one_rollout(rollout_id)\n\n            # Save checkpoint periodically\n            if should_run_periodic_action(\n                rollout_id, self.args.save_interval, self.num_rollout_per_epoch\n            ):\n                self.save_model(rollout_id)\n\n            # Calculate val loss periodically\n            if self.args.val_prompt_data and should_run_periodic_action(rollout_id, self.args.val_interval):\n                self.calculate_val_loss(rollout_id)\n\n        logger.info(f""[Rank {dist.get_rank()}] Training completed!"")\n\n\ndef set_sft_defaults(args: Namespace) -> Namespace:\n    """"""Set default values appropriate for SFT training.""""""\n    if not hasattr(args, ""loss_type"") or args.loss_type is None:\n        args.loss_type = ""sft_loss""\n\n    if not hasattr(args, ""advantage_estimator""):\n        args.advantage_estimator = None\n\n    args.offload_train = False\n    args.offload_rollout = False\n    args.colocate = False\n\n    return args\n\n\ndef main():\n    configure_logger()\n\n    args = parse_args()\n\n    args = set_sft_defaults(args)\n\n    trainer = SFTTrainer(args)\n    trainer.train()\n\n\nif __name__ == ""__main__"":\n    main()\n\n\n",python,selection_command
+58,2470130,"train_sft.py",29058,0,"",python,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-81de5df8-e118-4ce3-8f47-eb4a2db3766d1755539568535-2025_08_18-19.52.56.685/source.csv

@@ -0,0 +1,31 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"train_lam.py",0,0,"from dataclasses import dataclass, field\nimport os\nfrom typing import cast\n\nimport einops\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom models.lam import LatentActionModel\nfrom utils.dataloader import get_dataloader\nfrom utils.lr_utils import get_lr_schedule\nfrom utils.parameter_utils import count_parameters_by_component\n\njax.config.update(""jax_transfer_guard"", ""allow"")\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    batch_size: int = 36\n    vq_beta: float = 0.25\n    init_lr: float = 0.0\n    max_lr: float = 3e-5\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    warmup_steps: int = 5000\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    vq_reset_thresh: int = 50\n    # LAM\n    model_dim: int = 512\n    ffn_dim: int = 2048\n    latent_dim: int = 32\n    num_latents: int = 6\n    patch_size: int = 16\n    num_blocks: int = 4\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.0\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_lam""\n    tags: list[str] = field(default_factory=lambda: [""lam""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n    log_checkpoint_keep_period: int = 20000\n    wandb_id: str = """"\n    use_flash_attention: bool = True\n\n\nargs = tyro.cli(Args)\n\n\ndef lam_loss_fn(\n    model: LatentActionModel, inputs: dict\n) -> tuple[jax.Array, tuple[jax.Array, jax.Array, dict]]:\n    # --- Compute loss ---\n    gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n    inputs[""videos""] = gt.astype(args.dtype)\n    model.train()\n    outputs = model(inputs, training=True)\n    outputs[""recon""] = outputs[""recon""].astype(jnp.float32)\n    gt_future_frames = gt[:, 1:]\n    mse = jnp.square(gt_future_frames - outputs[""recon""]).mean()\n    q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n    commitment_loss = jnp.square(\n        outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n    ).mean()\n    loss = mse + q_loss + args.vq_beta * commitment_loss\n\n    # --- Compute validation metrics ---\n    gt = gt_future_frames.clip(0, 1).reshape(-1, *gt_future_frames.shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = jnp.asarray(pix.psnr(gt, recon)).mean()\n    ssim = jnp.asarray(pix.ssim(gt, recon)).mean()\n    count_fn = jax.vmap(lambda i: (outputs[""indices""] == i).sum())\n    index_counts = count_fn(jnp.arange(args.num_latents))\n    metrics = dict(\n        loss=loss,\n        mse=mse,\n        q_loss=q_loss,\n        commitment_loss=commitment_loss,\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage=(index_counts != 0).mean(),\n    )\n    return loss, (outputs[""recon""], index_counts, metrics)\n\n\n@nnx.jit\ndef train_step(\n    lam: LatentActionModel,\n    optimizer: nnx.Optimizer,\n    inputs: dict,\n    action_last_active: jax.Array,\n    rng: jax.Array,\n) -> tuple[jax.Array, jax.Array, jax.Array, dict]:\n    def loss_fn(\n        model: LatentActionModel,\n    ) -> tuple[jax.Array, tuple[jax.Array, jax.Array, dict]]:\n        return lam_loss_fn(model, inputs)\n\n    # --- Update model ---\n    (loss, (recon, idx_counts, metrics)), grads = nnx.value_and_grad(\n        loss_fn, has_aux=True\n    )(lam)\n    optimizer.update(grads)\n\n    # --- Reset inactive latent actions ---\n    codebook = lam.vq.codebook\n    num_codes = len(codebook)\n    active_codes = idx_counts != 0.0\n    action_last_active = jnp.where(active_codes, 0, action_last_active + 1)\n    p_code = active_codes / active_codes.sum()\n    reset_idxs = jax.random.choice(rng, num_codes, shape=(num_codes,), p=p_code)\n    do_reset = action_last_active >= args.vq_reset_thresh\n    new_codebook = jnp.where(\n        jnp.expand_dims(do_reset, -1), codebook[reset_idxs], codebook.value\n    )\n    lam.vq.codebook.value = new_codebook\n    action_last_active = jnp.where(do_reset, 0, action_last_active)\n    return loss, recon, action_last_active, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    lam = LatentActionModel(\n        in_dim=args.image_channels,\n        model_dim=args.model_dim,\n        ffn_dim=args.ffn_dim,\n        latent_dim=args.latent_dim,\n        num_latents=args.num_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.num_blocks,\n        num_heads=args.num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.codebook_dropout,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        rngs=rngs,\n    )\n\n    # Count parameters\n    _, params, _ = nnx.split(lam, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.param_dtype, # moments in full precision\n    )\n    optimizer = nnx.Optimizer(lam, tx)\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n    # --- Initialize checkpoint manager ---\n    step = 0\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n\n    # --- Create DataLoaderIterator from dataloader ---\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n\n    # --- Restore checkpoint ---\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        restored = checkpoint_manager.restore(\n            checkpoint_manager.latest_step(),\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),  # type: ignore\n            ),\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        grain_iterator = restored[""dataloader_state""]\n        step = checkpoint_manager.latest_step() or 0\n        print(f""Restored dataloader and model state from step {step}"")\n\n    # --- TRAIN LOOP ---\n    dataloader = (\n        jax.make_array_from_process_local_data(videos_sharding, elem)\n        for elem in grain_iterator\n    )\n    print(f""Starting training from step {step}..."")\n    action_last_active = jnp.zeros(args.num_latents, dtype=jnp.int32)\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            rng, _rng = jax.random.split(rng)\n\n            inputs = dict(videos=videos, rng=_rng)\n            rng, _rng = jax.random.split(rng)\n            loss, recon, action_last_active, metrics = train_step(\n                lam, optimizer, inputs, action_last_active, _rng\n            )\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0, 1:].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[0])),\n                            recon=wandb.Image(np.asarray(recon_seq[0])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            grain_iterator  # type: ignore\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n",python,tab
+2,1108,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"7:52:56 PM [info] Activating crowd-code\n7:52:56 PM [info] Recording started\n7:52:56 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,1287,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"7:52:57 PM [info] Git repository found\n7:52:57 PM [info] Git provider initialized successfully\n7:52:57 PM [info] Initial git state: [object Object]\n",Log,content
+4,5588,"train_lam.py",0,0,"",python,tab
+5,7587,"TERMINAL",0,0,"ssh",,terminal_focus
+6,11897,"TERMINAL",0,0,"bash",,terminal_focus
+7,92439,"train_lam.py",0,0,"",python,tab
+8,94124,"utils/nn.py",0,0,"import math\nfrom typing import Tuple, Callable, List\n\nfrom flax import nnx\nimport jax\nimport jax.numpy as jnp\nimport einops\n\n\nclass SpatioTemporalPositionalEncoding(nnx.Module):\n    """"""\n    Applies separate sinusoidal positional encodings to the temporal and spatial dimensions.\n    """"""\n    def __init__(self, d_model: int, max_len: int = 5000):\n        self.d_model = d_model\n        self.max_len = max_len\n\n        pe = jnp.zeros((self.max_len, self.d_model))\n        position = jnp.arange(0, self.max_len, dtype=jnp.float32)[:, None]\n        div_term = jnp.exp(\n            jnp.arange(0, self.d_model, 2) * (-math.log(10000.0) / self.d_model)\n        )\n        pe = pe.at[:, 0::2].set(jnp.sin(position * div_term))\n        pe = pe.at[:, 1::2].set(jnp.cos(position * div_term))\n        self.pe = nnx.Variable(pe)\n\n    def __call__(self, x: jax.Array) -> jax.Array:\n        """"""\n        Args:\n            x: The input tensor of shape (Batch, Time, Space, Dimension).\n\n        Returns:\n            The input tensor with positional encodings added.\n        """"""\n        assert x.ndim == 4, f""Input must be 4-dimensional, but got shape {x.shape}""\n\n        num_timesteps = x.shape[1]\n        num_spatial_patches = x.shape[2]\n\n        # Temporal positional encoding: (1, T, 1, D)\n        temporal_pe = self.pe.value[None, :num_timesteps, None, :]\n        x = x + temporal_pe\n\n        # Spatial positional encoding: (1, 1, S, D)\n        spatial_pe = self.pe.value[None, None, :num_spatial_patches, :]\n        x = x + spatial_pe\n\n        return x\n\n\nclass STBlock(nnx.Module):\n    def __init__(\n        self,\n        dim: int,\n        ffn_dim: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        rngs: nnx.Rngs,\n        sow_weights: bool,\n        sow_activations: bool,\n    ):\n        self.dim = dim\n        self.ffn_dim = ffn_dim\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.sow_weights = sow_weights\n        self.sow_activations = sow_activations\n\n        self.spatial_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n        self.spatial_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.dim,\n            qkv_features=self.dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=False\n            ),\n            rngs=rngs,\n            decode=False,\n        )\n\n        self.temporal_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n        self.temporal_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.dim,\n            qkv_features=self.dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=False,\n        )\n\n        self.ffn_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n        self.ffn_dense1 = nnx.Linear(\n            in_features=self.dim,\n            out_features=self.ffn_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense2 = nnx.Linear(\n            in_features=self.ffn_dim,\n            out_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    @nnx.remat\n    def __call__(self, x_BTNM: jax.Array) -> jax.Array:\n        # --- Spatial attention ---\n        z_BTNM = self.spatial_norm(x_BTNM)\n        z_BTNM = self.spatial_attention(z_BTNM, sow_weights=self.sow_weights)\n        x_BTNM = x_BTNM + z_BTNM\n\n        # --- Temporal attention ---\n        x_BNTM = x_BTNM.swapaxes(1, 2)\n        z_BNTM = self.temporal_norm(x_BNTM)\n        z_BNTM = self.temporal_attention(z_BNTM, sow_weights=self.sow_weights)\n        x_BNTM = x_BNTM + z_BNTM\n        x_BTNM = x_BNTM.swapaxes(1, 2)\n\n        # --- Feedforward ---\n        z_BTNM = self.ffn_norm(x_BTNM)\n        z_BTND = self.ffn_dense1(z_BTNM)\n        z_BTND = jax.nn.gelu(z_BTND)\n        z_BTNM = self.ffn_dense2(z_BTND)\n        x_BTNM = x_BTNM + z_BTNM\n        if self.sow_activations:\n            self.sow(nnx.Intermediate, 'activations', x_BTNM)\n        return x_BTNM\n\n\nclass STTransformer(nnx.Module):\n    """"""\n    Dimension keys:\n        B: batch size\n        T: number of frames\n        N: number of patches per frame\n        I: number of input features\n        M: model dimension\n        D: FFN dimension\n        V: vocabulary size\n    """"""\n    def __init__(\n        self,\n        input_dim: int,\n        model_dim: int,\n        ffn_dim: int,\n        out_dim: int,\n        num_blocks: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        rngs: nnx.Rngs,\n        sow_weights: bool = False,\n        sow_activations: bool = False,\n        sow_logits: bool = False,\n        max_len: int = 5000,\n    ):\n        self.input_dim = input_dim\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.out_dim = out_dim\n        self.num_blocks = num_blocks\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.sow_weights = sow_weights\n        self.sow_activations = sow_activations\n        self.sow_logits = sow_logits\n\n        self.input_norm1 = nnx.LayerNorm(\n            num_features=self.input_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n        self.input_dense = nnx.Linear(\n            in_features=self.input_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_norm2 = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n\n        self.pos_enc = SpatioTemporalPositionalEncoding(self.model_dim, max_len=max_len)\n\n        self.blocks = []\n        for _ in range(self.num_blocks):\n            self.blocks.append(\n                STBlock(\n                    dim=self.model_dim,\n                    ffn_dim=self.ffn_dim,\n                    num_heads=self.num_heads,\n                    dropout=self.dropout,\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                    use_flash_attention=self.use_flash_attention,\n                    rngs=rngs,\n                    sow_weights=self.sow_weights,\n                    sow_activations=self.sow_activations,\n                )\n            )\n\n        self.output_dense = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.out_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    def __call__(self, x_BTNI: jax.Array) -> jax.Array:\n        x_BTNI = self.input_norm1(x_BTNI)\n        x_BTNM = self.input_dense(x_BTNI)\n        x_BTNM = self.input_norm2(x_BTNM)\n        x_BTNM = self.pos_enc(x_BTNM)\n        for block in self.blocks:\n            x_BTNM = block(x_BTNM)\n\n        x_BTNV = self.output_dense(x_BTNM)\n        if self.sow_logits:\n            self.sow(nnx.Intermediate, 'logits', x_BTNV)\n        return x_BTNV\n\nclass TransformerBlock(nnx.Module):\n    def __init__(\n        self,\n        model_dim: int,\n        ffn_dim: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        decode: bool,\n        rngs: nnx.Rngs,\n        sow_weights: bool,\n        sow_activations: bool,\n    ):\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.decode = decode\n        self.sow_weights = sow_weights\n        self.sow_activations = sow_activations\n\n        self.temporal_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n        self.spatial_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n        self.ffn_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n        self.temporal_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.model_dim,\n            qkv_features=self.model_dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=self.decode,\n        )\n        self.spatial_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.model_dim,\n            qkv_features=self.model_dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=self.decode,\n        )\n        self.ffn_dense1 = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.ffn_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense2 = nnx.Linear(\n            in_features=self.ffn_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    # @nnx.remat\n    def __call__(self, x_BTNM: jax.Array, pos_index: Tuple[jax.Array, jax.Array] | None = None) -> jax.Array:\n        # --- Spatial attention ---\n        B, T, N, M = x_BTNM.shape\n        z_FNM = einops.rearrange(x_BTNM, ""b t n m -> (b t) n m"")\n        z_FNM = self.spatial_norm(z_FNM)\n        z_FNM = self.spatial_attention(z_FNM, sow_weights=self.sow_weights)\n        z_BTNM = einops.rearrange(z_FNM, ""(b t) n m -> b t n m"", t=T)\n        x_BTNM = x_BTNM + z_BTNM\n        # --- Temporal attention ---\n        z_PTM = einops.rearrange(x_BTNM, ""b t n m -> (b n) t m"")\n        z_PTM = self.temporal_norm(z_PTM)\n        z_PTM = self.temporal_attention(z_PTM, sow_weights=self.sow_weights)\n        z_BTNM = einops.rearrange(z_PTM, ""(b n) t m -> b t n m"", n=N)\n        x_BTNM = x_BTNM + z_BTNM\n        # --- Feedforward ---\n        z_BTNM = self.ffn_norm(x_BTNM)\n        z_BTND = self.ffn_dense1(z_BTNM)\n        z_BTND = jax.nn.gelu(z_BTND)\n        z_BTNM = self.ffn_dense2(z_BTND)\n        x_BTNM = x_BTNM + z_BTNM\n        if self.sow_activations:\n            self.sow(nnx.Intermediate, 'activations', x_BTNM)\n\n        return x_BTNM\n\nclass Transformer(nnx.Module):\n    """"""\n    Dimension keys:\n        B: batch size\n        T: number of frames\n        N: number of patches per frame\n        I: number of input features\n        M: model dimension\n        D: FFN dimension\n        V: vocabulary size\n        F: number of frames in batch\n        P: number of patch positions in batch\n    """"""\n    def __init__(\n        self,\n        input_dim: int,\n        model_dim: int,\n        ffn_dim: int,\n        out_dim: int,\n        num_blocks: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        decode: bool,\n        rngs: nnx.Rngs,\n        sow_logits: bool = False,\n        sow_weights: bool = False,\n        sow_activations: bool = False,\n        max_len: int = 5000,\n    ):\n        self.input_dim = input_dim\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.out_dim = out_dim\n        self.num_blocks = num_blocks\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.sow_logits = sow_logits\n        self.sow_weights = sow_weights\n        self.sow_activations = sow_activations\n\n        self.input_norm1 = nnx.LayerNorm(\n            num_features=self.input_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n        self.input_dense = nnx.Linear(\n            in_features=self.input_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_norm2 = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n\n        self.pos_enc = SpatioTemporalPositionalEncoding(self.model_dim, max_len=max_len)\n\n        self.blocks: List[TransformerBlock] = []\n        for _ in range(self.num_blocks):\n            self.blocks.append(\n                TransformerBlock(\n                    model_dim=self.model_dim,\n                    ffn_dim=self.ffn_dim,\n                    num_heads=self.num_heads,\n                    dropout=self.dropout,\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                    use_flash_attention=self.use_flash_attention,\n                    decode=decode,\n                    sow_weights=self.sow_weights,\n                    sow_activations=self.sow_activations,\n                    rngs=rngs,\n                )\n            )\n        self.output_dense = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.out_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    def __call__(self, x_BTNI: jax.Array, pos_index: Tuple[jax.Array, jax.Array] | None = None) -> jax.Array:\n        x_BTNI = self.input_norm1(x_BTNI)\n        x_BTNM = self.input_dense(x_BTNI)\n        x_BTNM = self.input_norm2(x_BTNM)\n        x_BTNM = self.pos_enc(x_BTNM)\n        for block in self.blocks:\n            x_BTNM = block(x_BTNM, pos_index)\n\n        x_BTNV = self.output_dense(x_BTNM)\n        if self.sow_logits:\n            self.sow(nnx.Intermediate, 'logits', x_BTNV)\n        return x_BTNV\n\ndef normalize(x: jax.Array) -> jax.Array:\n    return x / (jnp.linalg.norm(x, ord=2, axis=-1, keepdims=True) + 1e-8)\n\n\nclass VectorQuantizer(nnx.Module):\n    """"""\n    Dimension keys:\n        D: B * T * N\n        K: number of latents\n        L: latent dimension\n    """"""\n    def __init__(\n        self, latent_dim: int, num_latents: int, dropout: float, rngs: nnx.Rngs\n    ):\n        self.latent_dim = latent_dim\n        self.num_latents = num_latents\n        self.dropout = dropout\n\n        self.codebook = nnx.Param(\n            normalize(\n                nnx.initializers.lecun_uniform()(\n                    rngs.params(), (self.num_latents, self.latent_dim)\n                )\n            )\n        )\n        self.drop = nnx.Dropout(self.dropout, rngs=rngs)\n\n    def __call__(\n        self, x_DL: jax.Array, training: bool\n    ) -> Tuple[jax.Array, jax.Array, jax.Array, jax.Array]:\n        # --- Compute distances ---\n        x_DL = normalize(x_DL)\n        normalized_codebook_KL = normalize(self.codebook.value)\n        distance_DK = -jnp.matmul(x_DL, normalized_codebook_KL.T)\n        if training:\n            distance_DK = self.drop(distance_DK)\n\n        # --- Get indices and embeddings ---\n        indices_D = jnp.argmin(distance_DK, axis=-1)\n        z_DL = self.codebook[indices_D]\n\n        # --- Straight through estimator ---\n        z_q_DL = x_DL + jax.lax.stop_gradient(z_DL - x_DL)\n        return z_q_DL, z_DL, x_DL, indices_D\n\n    def get_codes(self, indices_E: jax.Array) -> jax.Array:\n        return self.codebook[indices_E]\n\n\ndef _create_flash_attention_fn(use_flash_attention: bool, is_causal: bool) -> Callable:\n    """"""\n    Create an attention function that uses flash attention if enabled.\n\n    flax.nnx.MultiHeadAttention provides tensors with shape (batch..., length, num_heads, head_dim),\n    but jax.nn.dot_product_attention expects (batch, length, num_heads, head_dim). We reshape to\n    ensure compatibility. cuDNN's flash attention additionally requires a sequence length that\n    is a multiple of 4. We pad the sequence length to the nearest multiple of 4 and mask\n    accordingly. Note that cuDNN requires the mask to be broadcast before calling the attention\n    function due to strict shape checking.\n    """"""\n\n    def attention_fn(query_BTHD, key_BSHD, value_BSHD, bias=None, mask_B111=None, **kwargs):\n        implementation = ""cudnn"" if use_flash_attention else None\n\n        def _merge_batch_dims(x):\n            return einops.rearrange(x, ""... l h k -> (...) l h k"")\n\n        def _pad(x, pad_size):\n            return jnp.pad(x, ((0, 0), (0, pad_size), (0, 0), (0, 0)))\n\n        original_shape = query_BTHD.shape\n        T = query_BTHD.shape[-3]\n        S = key_BSHD.shape[-3]\n\n        # Pad to nearest multiple of 4\n        Q = ((T + 3) // 4) * 4\n        pad_size_Q = Q - T\n        K = ((S + 3) // 4) * 4\n        pad_size_K = K - S\n\n        query_BQHD = _pad(_merge_batch_dims(query_BTHD), pad_size_Q)\n        key_BKHD = _pad(_merge_batch_dims(key_BSHD), pad_size_K)\n        value_BKHD = _pad(_merge_batch_dims(value_BSHD), pad_size_K)\n\n        attention_mask = jnp.ones((Q, K), dtype=jnp.bool_)\n        attention_mask = attention_mask.at[T:, :].set(False)\n        attention_mask = attention_mask.at[:, S:].set(False)\n\n        mask_11TS = attention_mask[jnp.newaxis, jnp.newaxis, :, :]\n\n        bias_4d = jnp.pad(_merge_batch_dims(bias), ((0, 0), (0, 0), (0, pad_size_Q), (0, pad_size_K))) if bias is not None else None\n\n        # NOTE: jax.nn.dot_product_attention does not support dropout\n        output_4d = jax.nn.dot_product_attention(\n            query=query_BQHD,\n            key=key_BKHD,\n            value=value_BKHD,\n            bias=bias_4d,\n            mask=mask_11TS,\n            implementation=implementation,\n            is_causal=is_causal,\n        )\n        return output_4d[..., :T, :, :].reshape(original_shape)\n\n    return attention_fn\n",python,tab
+9,94459,"train_lam.py",0,0,"",python,tab
+10,96774,"test/test_nan.ipynb",0,0,"%pwd",python,tab
+11,150183,"test/test_nan.ipynb",0,0,"",python,tab
+12,247542,"test/test_nan.ipynb",0,0,"# Restore a dynamics checkpoint and enable sowing\nimport os\nfrom typing import Dict\n\nimport jax\nimport jax.numpy as jnp\nimport flax.nnx as nnx\nimport optax\nimport orbax.checkpoint as ocp\nimport grain\n\nfrom utils.dataloader import get_dataloader\nfrom models.lam import LatentActionModel\n\n# Adjust to your checkpoint directory, dataset directory, and dynamics type\nckpt_dir = ""/fast/project/HFMI_SynergyUnit/jafar_ws/checkpoints/coinrun/lam/train_lam_coinrun_reproduction_20067/100000_ckpt""\ndata_dir = ""/fast/project/HFMI_SynergyUnit/jafar_ws/data/coinrun/array_records_10m""\nnum_steps: int = 200_000\nseed: int = 0\nseq_len: int = 16\nimage_channels: int = 3\nimage_height: int = 64\nimage_width: int = 64\nsave_ckpt: bool = False\nrestore_ckpt: bool = False\n# Optimization\nbatch_size: int = 36\nvq_beta: float = 0.25\ninit_lr: float = 0.0\nmax_lr: float = 3e-5\ndecay_end: float = 0.0\nwsd_decay_steps: int = (\n    10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n)\nwarmup_steps: int = 5000\nlr_schedule: str = ""wsd""  # supported options: wsd, cos\nvq_reset_thresh: int = 50\n# LAM\nmodel_dim: int = 512\nffn_dim: int = 2048\nlatent_dim: int = 32\nnum_latents: int = 6\npatch_size: int = 16\nnum_blocks: int = 4\nnum_heads: int = 8\ndropout: float = 0.0\ncodebook_dropout: float = 0.0\nparam_dtype = jnp.float32\ndtype = jnp.bfloat16\n# Logging\nlog_interval: int = 5\nlog_image_interval: int = 250\nuse_flash_attention: bool = True\n\n# Build model graph matching the checkpoint\nrng = jax.random.key(seed)\nrng, _rng = jax.random.split(rng)\nrngs = nnx.Rngs(_rng)\nlam = LatentActionModel(\n    in_dim=image_channels,\n    model_dim=model_dim,\n    ffn_dim=ffn_dim,\n    latent_dim=latent_dim,\n    num_latents=num_latents,\n    patch_size=patch_size,\n    num_blocks=num_blocks,\n    num_heads=num_heads,\n    dropout=dropout,\n    codebook_dropout=codebook_dropout,\n    param_dtype=param_dtype,\n    dtype=dtype,\n    use_flash_attention=use_flash_attention,\n    rngs=rngs,\n)\n\n# Optimizer (matches training opt hyperparams; lr value is irrelevant for restore)\ntx = optax.adamw(\n    learning_rate=max_lr,\n    b1=0.9,\n    b2=0.9,\n    weight_decay=1e-4,\n    mu_dtype=param_dtype,\n)\noptimizer = nnx.Optimizer(lam, tx)\n",python,tab
+13,247543,"test/test_nan.ipynb",483,0,"",python,selection_mouse
+14,247662,"test/test_nan.ipynb",476,11,"100000_ckpt",python,selection_mouse
+15,249645,"test/test_nan.ipynb",483,0,"",python,selection_mouse
+16,250077,"test/test_nan.ipynb",486,0,"",python,selection_command
+17,250338,"test/test_nan.ipynb",486,1,"t",python,selection_command
+18,250487,"test/test_nan.ipynb",476,11,"100000_ckpt",python,selection_command
+19,250935,"test/test_nan.ipynb",363,124,"ckpt_dir = ""/fast/project/HFMI_SynergyUnit/jafar_ws/checkpoints/coinrun/lam/train_lam_coinrun_reproduction_20067/100000_ckpt",python,selection_command
+20,251750,"test/test_nan.ipynb",372,115,"= ""/fast/project/HFMI_SynergyUnit/jafar_ws/checkpoints/coinrun/lam/train_lam_coinrun_reproduction_20067/100000_ckpt",python,selection_command
+21,251922,"test/test_nan.ipynb",374,113,"""/fast/project/HFMI_SynergyUnit/jafar_ws/checkpoints/coinrun/lam/train_lam_coinrun_reproduction_20067/100000_ckpt",python,selection_command
+22,252426,"test/test_nan.ipynb",375,112,"/fast/project/HFMI_SynergyUnit/jafar_ws/checkpoints/coinrun/lam/train_lam_coinrun_reproduction_20067/100000_ckpt",python,selection_command
+23,252821,"test/test_nan.ipynb",375,0,"",python,selection_command
+24,253714,"test/test_nan.ipynb",0,0,"",python,tab
+25,256448,"TERMINAL",0,0,"",,terminal_focus
+26,256910,"TERMINAL",0,0,"source /home/franz.srambical/jafar/.venv/bin/activate",,terminal_command
+27,258932,"TERMINAL",0,0,"ls /fast/project/HFMI_SynergyUnit/jafar_ws/checkpoints/coinrun/lam/train_lam_coinrun_reproduction_20067/100000_ckpt",,terminal_command
+28,258951,"TERMINAL",0,0,"]633;C100000\r\n]0;franz.srambical@hai-login2:~/jafar",,terminal_output
+29,263891,"TERMINAL",0,0,"ssh",,terminal_focus
+30,265268,"TERMINAL",0,0,"bash",,terminal_focus

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-99d0e110-10e9-436d-a28e-33d9ed7c63831756230556162-2025_08_26-19.49.24.278/source.csv

@@ -0,0 +1,24 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,4,"train_tokenizer.py",0,0,"import os\n\n# os.environ['XLA_FLAGS'] = (\n#     '--xla_python_client_mem_fraction=.98 '\n# )\n# FIXME (f.srambical): test whether this increases throughput\n# os.environ['XLA_FLAGS'] = (\n#     '--xla_gpu_enable_latency_hiding_scheduler=true '\n#     '--xla_gpu_enable_async_collectives=true '\n# )\n\nfrom dataclasses import dataclass, field\nfrom typing import cast, Optional\n\nimport einops\nimport itertools\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom models.tokenizer import TokenizerVQVAE\nfrom utils.dataloader import get_dataloader\nfrom utils.train_utils import (\n    get_lr_schedule,\n    count_parameters_by_component,\n    print_mem_stats,\n    print_compiled_memory_stats,\n    print_compiled_cost_analysis,\n)\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 300_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    vq_beta: float = 0.25\n    batch_size: int = 48\n    init_lr: float = 0.0\n    max_lr: float = 3e-4\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        20000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    warmup_steps: int = 10000\n    # Tokenizer\n    model_dim: int = 512\n    ffn_dim: int = 2048\n    latent_dim: int = 32\n    num_latents: int = 1024\n    patch_size: int = 4\n    num_blocks: int = 4\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.01\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_tokenizer""\n    tags: list[str] = field(default_factory=lambda: [""tokenizer""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n    log_checkpoint_keep_period: int = 20000\n    log_gradients: bool = False\n    wandb_id: str = """"\n    use_flash_attention: bool = True\n\n\ndef build_model(args: Args, rng: jax.Array) -> tuple[TokenizerVQVAE, jax.Array]:\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    return (\n        TokenizerVQVAE(\n            in_dim=args.image_channels,\n            model_dim=args.model_dim,\n            ffn_dim=args.ffn_dim,\n            latent_dim=args.latent_dim,\n            num_latents=args.num_latents,\n            patch_size=args.patch_size,\n            num_blocks=args.num_blocks,\n            num_heads=args.num_heads,\n            dropout=args.dropout,\n            codebook_dropout=args.codebook_dropout,\n            param_dtype=args.param_dtype,\n            dtype=args.dtype,\n            use_flash_attention=args.use_flash_attention,\n            rngs=rngs,\n        ),\n        rng,\n    )\n\n\ndef build_optimizer(\n    model: TokenizerVQVAE, args: Args\n) -> tuple[nnx.Optimizer, optax.Schedule]:\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.param_dtype,  # moments in full precision\n    )\n    optimizer = nnx.Optimizer(model, tx)\n    return optimizer, lr_schedule\n\n\ndef build_mesh_and_sharding(\n    num_devices: int,\n) -> tuple[Mesh, NamedSharding, NamedSharding]:\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n    return mesh, replicated_sharding, videos_sharding\n\n\ndef shard_optimizer_states(\n    optimizer: nnx.Optimizer, replicated_sharding: NamedSharding\n) -> None:\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n\ndef build_dataloader(args: Args) -> grain.DataLoaderIterator:\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n    return grain_iterator\n\n\ndef build_checkpoint_manager(args: Args) -> ocp.CheckpointManager:\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n    return checkpoint_manager\n\n\ndef restore_checkpoint_if_needed(\n    args: Args,\n    checkpoint_manager: ocp.CheckpointManager,\n    optimizer: nnx.Optimizer,\n    grain_iterator: grain.DataLoaderIterator,\n    restore_step: Optional[int] = None,\n) -> tuple[int, nnx.Optimizer, grain.DataLoaderIterator]:\n    step = 0\n    if restore_step is None:\n        restore_step = checkpoint_manager.latest_step()\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        restored = checkpoint_manager.restore(\n            restore_step,\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),  # type: ignore\n            ),\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        grain_iterator = restored[""dataloader_state""]\n        step = restore_step or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    return step, optimizer, grain_iterator\n\n\ndef main(args: Args) -> None:\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    tokenizer, rng = build_model(args, rng)\n\n    _, params, _ = nnx.split(tokenizer, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    optimizer, lr_schedule = build_optimizer(tokenizer, args)\n    del tokenizer\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    mesh, replicated_sharding, videos_sharding = build_mesh_and_sharding(num_devices)\n\n    shard_optimizer_states(optimizer, replicated_sharding)\n\n    # --- Initialize checkpoint manager ---\n    checkpoint_manager = build_checkpoint_manager(args)\n\n    # --- Create DataLoaderIterator from dataloader ---\n    grain_iterator = build_dataloader(args)\n\n    # --- Restore checkpoint ---\n    step, optimizer, grain_iterator = restore_checkpoint_if_needed(\n        args, checkpoint_manager, optimizer, grain_iterator\n    )\n\n    # --- Define loss and train step (close over args) ---\n    def tokenizer_loss_fn(\n        model: TokenizerVQVAE, inputs: dict\n    ) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n        gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n        inputs[""videos""] = gt.astype(args.dtype)\n        model.train()\n        outputs = model(inputs, training=True)\n        outputs[""recon""] = outputs[""recon""].astype(jnp.float32)\n        mse = jnp.square(gt - outputs[""recon""]).mean()\n        q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n        commitment_loss = jnp.square(\n            outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n        ).mean()\n        loss = mse + q_loss + args.vq_beta * commitment_loss\n\n        gt_clipped = gt.clip(0, 1).reshape(-1, *gt.shape[2:])\n        recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n        psnr = jnp.asarray(pix.psnr(gt_clipped, recon)).mean()\n        ssim = jnp.asarray(pix.ssim(gt_clipped, recon)).mean()\n        _, index_counts = jnp.unique_counts(\n            jnp.ravel(outputs[""indices""]), size=args.num_latents, fill_value=0\n        )\n        codebook_usage = (index_counts != 0).mean()\n        metrics = dict(\n            loss=loss,\n            mse=mse,\n            q_loss=q_loss,\n            commitment_loss=commitment_loss,\n            psnr=psnr,\n            ssim=ssim,\n            codebook_usage=codebook_usage,\n        )\n        return loss, (outputs[""recon""], metrics)\n\n    @nnx.jit(donate_argnums=0)\n    def train_step(\n        optimizer: nnx.Optimizer, inputs: dict\n    ) -> tuple[jax.Array, jax.Array, dict]:\n        def loss_fn(model: TokenizerVQVAE) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n            return tokenizer_loss_fn(model, inputs)\n\n        (loss, (recon, metrics)), grads = nnx.value_and_grad(loss_fn, has_aux=True)(\n            optimizer.model\n        )\n        optimizer.update(grads)\n        if args.log_gradients:\n            metrics[""encoder_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""encoder""]\n            )\n            metrics[""vq_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""vq""]\n            )\n            metrics[""decoder_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""decoder""]\n            )\n        return loss, recon, metrics\n\n    # --- TRAIN LOOP ---\n    dataloader = (\n        jax.make_array_from_process_local_data(videos_sharding, elem)\n        for elem in grain_iterator\n    )\n    if jax.process_index() == 0:\n        first_videos = next(dataloader)\n        sample_inputs = dict(videos=first_videos)\n        compiled = train_step.lower(optimizer, sample_inputs).compile()\n        print_compiled_memory_stats(compiled.memory_analysis())\n        print_compiled_cost_analysis(compiled.cost_analysis())\n        # Do not skip the first batch during training\n        dataloader = itertools.chain([first_videos], dataloader)\n    print(f""Starting training from step {step}..."")\n    first_step = step\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            inputs = dict(videos=videos)\n            loss, recon, metrics = train_step(optimizer, inputs)\n            if step == first_step:\n                print_mem_stats(""After params initialized"")\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[0])),\n                            recon=wandb.Image(np.asarray(recon_seq[0])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            grain_iterator  # type: ignore\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(Args)\n    main(args)\n",python,tab
+2,146,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"7:49:24 PM [info] Activating crowd-code\n7:49:24 PM [info] Recording started\n7:49:24 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,167,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"7:49:24 PM [info] Git repository found\n7:49:24 PM [info] Git provider initialized successfully\n7:49:24 PM [info] Initial git state: [object Object]\n",Log,content
+4,1210,"train_tokenizer.py",0,0,"",python,tab
+5,3346,"train_tokenizer.py",293,0,"",python,selection_command
+6,4170,"train_tokenizer.py",292,0,"",python,selection_command
+7,4174,"train_tokenizer.py",293,0,"",python,selection_command
+8,4247,"train_tokenizer.py",292,0,"",python,selection_command
+9,10075,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+10,12944,"train_tokenizer.py",0,0,"",python,tab
+11,12946,"TERMINAL",0,0,"",,terminal_focus
+12,13563,"TERMINAL",0,0,"source /home/franz.srambical/jafar/.venv/bin/activate",,terminal_command
+13,13566,"TERMINAL",0,0,"]633;C]0;franz.srambical@hai-login2:~/jafar",,terminal_output
+14,14811,"TERMINAL",0,0,"salloc --gpus=1 --ntasks-per-node=1 --cpus-per-task=1 --mem=100G",,terminal_command
+15,14878,"TERMINAL",0,0,"]633;Csalloc: Pending job allocation 23994\r\nsalloc: job 23994 queued and waiting for resources\r\n",,terminal_output
+16,20973,"TERMINAL",0,0,"",,terminal_focus
+17,21466,"TERMINAL",0,0,"source /home/franz.srambical/jafar/.venv/bin/activate",,terminal_command
+18,22695,"TERMINAL",0,0,"squeue",,terminal_command
+19,22706,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  23994 franz.sram interacti     1     1 PD 2025-08-26T19:49:39                 N/A     0:00 1-00:00:00 (Nodes required for job are DOWN, DRAINED or reserved for jobs in higher priority partitions)\r\n  23858   xiao.liu interacti     1    64  R 2025-08-26T10:11:29 2025-08-26T10:11:29  9:38:17   23:59:00 hai003\r\n  23857   xiao.liu interacti     1    64  R 2025-08-26T10:11:23 2025-08-26T10:11:23  9:38:23   23:59:00 hai002\r\n  23856   xiao.liu interacti     1    64  R 2025-08-26T10:10:41 2025-08-26T10:10:41  9:39:05   23:59:00 hai004\r\n  23941 yoland.sav  standard     1   100 PD 2025-08-26T15:28:50 2025-08-26T23:47:01     0:00 1-00:00:00 (Priority)\r\n23921_[ yoland.sav  standard     1    28 PD 2025-08-26T14:59:53 2025-08-26T21:13:29     0:00    4:00:00 (Resources)\r\n23921_3 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T19:48:55     0:51    4:00:00 hai004\r\n23921_3 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T19:47:01     2:45    4:00:00 hai008\r\n23921_3 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T19:30:25    19:21    4:00:00 hai008\r\n23921_3 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T19:00:30    49:16    4:00:00 hai005\r\n23921_3 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T19:00:30    49:16    4:00:00 hai007\r\n23921_3 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T19:00:30    49:16    4:00:00 hai008\r\n23921_3 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T19:00:30    49:16    4:00:00 hai003\r\n23921_3 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T19:00:30    49:16    4:00:00 hai006\r\n23921_3 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T19:00:25    49:21    4:00:00 hai002\r\n23921_2 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T18:29:15  1:20:31    4:00:00 hai007\r\n23921_2 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T18:26:15  1:23:31    4:00:00 hai002\r\n23921_2 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T18:02:45  1:47:01    4:00:00 hai007\r\n23921_2 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T17:50:27  1:59:19    4:00:00 hai006\r\n23921_2 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T17:43:21  2:06:25    4:00:00 hai005\r\n23921_2 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T17:33:53  2:15:53    4:00:00 hai004\r\n23921_2 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T17:17:10  2:32:36    4:00:00 hai006\r\n23921_2 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T17:16:45  2:33:01    4:00:00 hai003\r\n23921_1 yoland.sav  standard     1    56  R 2025-08-26T14:59:53 2025-08-26T17:13:29  2:36:17    4:00:00 hai005\r\n  23855 yoland.sav  standard     1   128  R 2025-08-26T09:14:30 2025-08-26T09:14:31 10:35:15 1-00:00:00 hai001\r\n]0;franz.srambical@hai-login2:~/jafar",,terminal_output
+20,36826,"TERMINAL",0,0,"salloc",,terminal_focus
+21,37949,"TERMINAL",0,0,"bash",,terminal_focus
+22,66139,"TERMINAL",0,0,"salloc",,terminal_focus
+23,77826,"TERMINAL",0,0,"^Csalloc: Job allocation 23994 has been revoked.\r\nsalloc: Job aborted due to signal\r\n]0;franz.srambical@hai-login2:~/jafar",,terminal_output

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-9a63c010-bab0-4339-9cb8-bb9e93007fed1759485346203-2025_10_03-11.55.53.231/source.csv

@@ -0,0 +1,17 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"jasmine/utils/dataloader.py",0,0,"import jax\nimport numpy as np\nimport grain\nfrom typing import Any\nimport pickle\n\n\nclass EpisodeLengthFilter(grain.transforms.Filter):\n    """"""\n    A Grain Filter that keeps only episodes with sufficient length.\n    """"""\n\n    def __init__(self, seq_len: int, image_h: int, image_w: int, image_c: int):\n        """"""Initializes the filter with sequence length requirements.""""""\n        self.seq_len = seq_len\n        self.image_h = image_h\n        self.image_w = image_w\n        self.image_c = image_c\n\n    def filter(self, element: Any) -> bool:\n        """"""\n        Filters episodes based on length.\n\n        Args:\n            element: A dictionary representing one record from the DataSource.\n                     Expected to contain 'raw_video' (bytes) and 'sequence_length' (int)\n\n        Returns:\n            True if the episode has sufficient length, False otherwise.\n        """"""\n        assert isinstance(element, bytes)\n        element = pickle.loads(element)\n\n        current_episode_len = element[""sequence_length""]\n        if current_episode_len < self.seq_len:\n            print(\n                f""Filtering out episode with length {current_episode_len}, which is ""\n                f""shorter than the requested sequence length {self.seq_len}.""\n            )\n            return False\n\n        return True\n\n\nclass ProcessEpisodeAndSlice(grain.transforms.RandomMap):\n    """"""\n    A Grain Transformation that combines parsing, slicing, and normalizing.\n    """"""\n\n    def __init__(self, seq_len: int, image_h: int, image_w: int, image_c: int):\n        """"""Initializes the transformation with processing parameters.""""""\n        self.seq_len = seq_len\n        self.image_h = image_h\n        self.image_w = image_w\n        self.image_c = image_c\n\n    def random_map(self, element: dict, rng: np.random.Generator) -> Any:\n        """"""\n        Processes a single raw episode from the data source.\n\n        Args:\n            element: A dictionary representing one record from the DataSource.\n                     Expected to contain 'raw_video' (bytes) and 'sequence_length' (int)\n            rng: A per-record random number generator provided by the Grain sampler.\n\n        Returns:\n            A processed video sequence as a NumPy array with shape\n            (seq_len, height, width, channels) and dtype float32.\n        """"""\n        assert isinstance(element, bytes)\n        element = pickle.loads(element)\n\n        video_shape = (\n            element[""sequence_length""],\n            self.image_h,\n            self.image_w,\n            self.image_c,\n        )\n        episode_tensor = np.frombuffer(element[""raw_video""], dtype=np.uint8)\n        episode_tensor = episode_tensor.reshape(video_shape)\n\n        current_episode_len = episode_tensor.shape[0]\n        if current_episode_len < self.seq_len:\n            raise ValueError(\n                f""Episode length {current_episode_len} is shorter than ""\n                f""requested sequence length {self.seq_len}. This should ""\n                f""have been filtered out.""\n            )\n\n        max_start_idx = current_episode_len - self.seq_len\n\n        start_idx = rng.integers(0, max_start_idx + 1)\n\n        seq = episode_tensor[start_idx : start_idx + self.seq_len]\n\n        data_dict = {""videos"": seq}\n        if ""actions"" in element.keys():\n            actions_tensor = np.array(element[""actions""])\n            actions = actions_tensor[start_idx : start_idx + self.seq_len]\n            data_dict[""actions""] = actions\n\n        return data_dict\n\n\ndef get_dataloader(\n    array_record_paths: list[str],\n    seq_len: int,\n    global_batch_size: int,\n    image_h: int,\n    image_w: int,\n    image_c: int,\n    num_workers: int = 1,\n    prefetch_buffer_size: int = 1,\n    seed: int = 42,\n):\n    """"""\n    Creates a data loading pipeline using Grain.\n    """"""\n    if not array_record_paths:\n        raise ValueError(""array_record_paths list cannot be empty."")\n\n    num_processes = jax.process_count()\n\n    if global_batch_size % num_processes != 0:\n        raise ValueError(\n            f""Global batch size {global_batch_size} must be divisible by ""\n            f""the number of JAX processes {num_processes} for proper sharding.""\n        )\n    per_process_batch_size = global_batch_size // num_processes\n\n    source = grain.sources.ArrayRecordDataSource(array_record_paths)\n\n    sampler = grain.samplers.IndexSampler(\n        num_records=len(source),\n        shard_options=grain.sharding.ShardByJaxProcess(drop_remainder=True),\n        shuffle=True,\n        num_epochs=None,\n        seed=seed,\n    )\n\n    operations = [\n        EpisodeLengthFilter(\n            seq_len=seq_len, image_h=image_h, image_w=image_w, image_c=image_c\n        ),\n        ProcessEpisodeAndSlice(\n            seq_len=seq_len, image_h=image_h, image_w=image_w, image_c=image_c\n        ),\n        grain.transforms.Batch(batch_size=per_process_batch_size, drop_remainder=True),\n    ]\n\n    read_options = grain.ReadOptions(\n        prefetch_buffer_size=prefetch_buffer_size,\n        num_threads=1,\n    )\n    dataloader = grain.DataLoader(\n        data_source=source,\n        sampler=sampler,\n        operations=operations,\n        worker_count=num_workers,\n        worker_buffer_size=1,\n        read_options=read_options,\n    )\n\n    return dataloader\n",python,tab
+2,125,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"11:55:53 AM [info] Activating crowd-code\n11:55:53 AM [info] Recording started\n11:55:53 AM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,239,"TERMINAL",0,0,"",,terminal_focus
+4,240,"jasmine/utils/dataloader.py",0,0,"",python,tab
+5,240,"TERMINAL",0,0,"",,terminal_focus
+6,240,"TERMINAL",0,0,"",,terminal_focus
+7,400,"TERMINAL",0,0,"source /home/franz.srambical/jafar/.venv/bin/activate",,terminal_command
+8,400,"TERMINAL",0,0,"source /home/franz.srambical/jafar/.venv/bin/activate",,terminal_command
+9,4321,"TERMINAL",0,0,"",,terminal_command
+10,11087,"TERMINAL",0,0,"",,terminal_command
+11,119391,"TERMINAL",0,0,"",,terminal_command
+12,598867,"jasmine/train_tokenizer.py",0,0,"import os\n\nos.environ.setdefault(""XLA_PYTHON_CLIENT_MEM_FRACTION"", ""0.98"")\n\nfrom dataclasses import dataclass, field\nfrom typing import cast, Optional\n\nimport einops\nimport itertools\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom models.tokenizer import TokenizerVQVAE\nfrom utils.dataloader import get_dataloader\nfrom utils.train_utils import (\n    get_lr_schedule,\n    count_parameters_by_component,\n    print_mem_stats,\n    print_compiled_memory_stats,\n    print_compiled_cost_analysis,\n)\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 300_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 64\n    image_width: int = 64\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    vq_beta: float = 0.25\n    batch_size: int = 48\n    init_lr: float = 0.0\n    max_lr: float = 3e-4\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        30_000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    warmup_steps: int = 10000\n    # Tokenizer\n    model_dim: int = 512\n    ffn_dim: int = 2048\n    latent_dim: int = 32\n    num_latents: int = 1024\n    patch_size: int = 16\n    num_blocks: int = 4\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.01\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    use_flash_attention: bool = True\n    # Logging\n    log: bool = True\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_tokenizer""\n    tags: list[str] = field(default_factory=lambda: [""tokenizer""])\n    log_interval: int = 50\n    log_image_interval: int = 1000\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 1000\n    log_checkpoint_keep_period: int = 20_000\n    log_gradients: bool = False\n    val_data_dir: str = """"\n    val_interval: int = 20_000\n    val_steps: int = 50\n    wandb_id: str = """"\n\n\ndef build_model(args: Args, rng: jax.Array) -> tuple[TokenizerVQVAE, jax.Array]:\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    return (\n        TokenizerVQVAE(\n            in_dim=args.image_channels,\n            model_dim=args.model_dim,\n            ffn_dim=args.ffn_dim,\n            latent_dim=args.latent_dim,\n            num_latents=args.num_latents,\n            patch_size=args.patch_size,\n            num_blocks=args.num_blocks,\n            num_heads=args.num_heads,\n            dropout=args.dropout,\n            codebook_dropout=args.codebook_dropout,\n            param_dtype=args.param_dtype,\n            dtype=args.dtype,\n            use_flash_attention=args.use_flash_attention,\n            rngs=rngs,\n        ),\n        rng,\n    )\n\n\ndef build_optimizer(model: TokenizerVQVAE, args: Args) -> nnx.ModelAndOptimizer:\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.param_dtype,  # moments in full precision\n    )\n    optimizer = nnx.ModelAndOptimizer(model, tx)\n    return optimizer\n\n\ndef build_mesh_and_sharding(\n    num_devices: int,\n) -> tuple[Mesh, NamedSharding, NamedSharding]:\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n    return mesh, replicated_sharding, videos_sharding\n\n\ndef shard_optimizer_states(\n    optimizer: nnx.ModelAndOptimizer, replicated_sharding: NamedSharding\n) -> None:\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n\ndef build_dataloader(args: Args, data_dir: str) -> grain.DataLoaderIterator:\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(data_dir, x)\n        for x in os.listdir(data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=2,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n    return grain_iterator\n\n\ndef build_checkpoint_manager(args: Args) -> Optional[ocp.CheckpointManager]:\n    if args.restore_ckpt or args.save_ckpt:\n        handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n        handler_registry.add(\n            ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n        )\n        handler_registry.add(\n            ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n        )\n        handler_registry.add(\n            ""train_dataloader_state"",\n            grain.checkpoint.CheckpointSave,\n            cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n        )\n        handler_registry.add(\n            ""train_dataloader_state"",\n            grain.checkpoint.CheckpointRestore,\n            cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n        )\n        if args.val_data_dir:\n            handler_registry.add(\n                ""val_dataloader_state"",\n                grain.checkpoint.CheckpointSave,\n                cast(\n                    ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler\n                ),\n            )\n            handler_registry.add(\n                ""val_dataloader_state"",\n                grain.checkpoint.CheckpointRestore,\n                cast(\n                    ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler\n                ),\n            )\n        checkpoint_options = ocp.CheckpointManagerOptions(\n            save_interval_steps=args.log_checkpoint_interval,\n            max_to_keep=3,\n            keep_period=args.log_checkpoint_keep_period,\n            step_format_fixed_length=6,\n            cleanup_tmp_directories=True,\n        )\n        checkpoint_manager = ocp.CheckpointManager(\n            args.ckpt_dir,\n            options=checkpoint_options,\n            handler_registry=handler_registry,\n        )\n        return checkpoint_manager\n    else:\n        return None\n\n\ndef restore_checkpoint_if_needed(\n    args: Args,\n    checkpoint_manager: Optional[ocp.CheckpointManager],\n    optimizer: nnx.ModelAndOptimizer,\n    train_iterator: grain.DataLoaderIterator,\n    val_iterator: Optional[grain.DataLoaderIterator],\n    restore_step: Optional[int] = None,\n) -> tuple[\n    int, nnx.ModelAndOptimizer, grain.DataLoaderIterator, grain.DataLoaderIterator\n]:\n    step = 0\n    if checkpoint_manager and restore_step is None:\n        restore_step = checkpoint_manager.latest_step()\n    if args.restore_ckpt:\n        assert checkpoint_manager is not None\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        if val_iterator:\n            restore_args = ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                train_dataloader_state=grain.checkpoint.CheckpointRestore(train_iterator),  # type: ignore\n                val_dataloader_state=grain.checkpoint.CheckpointRestore(val_iterator),  # type: ignore\n            )\n        else:\n            restore_args = ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                train_dataloader_state=grain.checkpoint.CheckpointRestore(train_iterator),  # type: ignore\n            )\n        restored = checkpoint_manager.restore(restore_step, args=restore_args)\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        train_iterator = restored[""train_dataloader_state""]\n        if val_iterator:\n            val_iterator = restored[""val_dataloader_state""]\n        step = restore_step or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    return step, optimizer, train_iterator, val_iterator\n\n\ndef main(args: Args) -> None:\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    tokenizer, rng = build_model(args, rng)\n\n    _, params, _ = nnx.split(tokenizer, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    optimizer = build_optimizer(tokenizer, args)\n    del tokenizer\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    _, replicated_sharding, videos_sharding = build_mesh_and_sharding(num_devices)\n\n    shard_optimizer_states(optimizer, replicated_sharding)\n\n    # --- Initialize checkpoint manager ---\n    checkpoint_manager = build_checkpoint_manager(args)\n\n    # --- Create DataLoaderIterator from dataloader ---\n    train_iterator = build_dataloader(args, args.data_dir)\n    val_iterator = None\n    if args.val_data_dir:\n        val_iterator = build_dataloader(args, args.val_data_dir)\n\n    # --- Restore checkpoint ---\n    step, optimizer, train_iterator, val_iterator = restore_checkpoint_if_needed(\n        args, checkpoint_manager, optimizer, train_iterator, val_iterator\n    )\n\n    # --- Define loss and train step (close over args) ---\n    def tokenizer_loss_fn(\n        model: TokenizerVQVAE, inputs: dict, training: bool = False\n    ) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n        gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n        inputs[""videos""] = gt.astype(args.dtype)\n        outputs = model(inputs, training=training)\n        outputs[""recon""] = outputs[""recon""].astype(jnp.float32)\n        mse = jnp.square(gt - outputs[""recon""]).mean()\n        q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n        commitment_loss = jnp.square(\n            outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n        ).mean()\n        loss = mse + q_loss + args.vq_beta * commitment_loss\n\n        gt_clipped = gt.clip(0, 1).reshape(-1, *gt.shape[2:])\n        recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n        psnr = jnp.asarray(pix.psnr(gt_clipped, recon)).mean()\n        ssim = jnp.asarray(pix.ssim(gt_clipped, recon)).mean()\n        _, index_counts = jnp.unique_counts(\n            jnp.ravel(outputs[""indices""]), size=args.num_latents, fill_value=0\n        )\n        codebook_usage = (index_counts != 0).mean()\n        metrics = dict(\n            loss=loss,\n            mse=mse,\n            q_loss=q_loss,\n            commitment_loss=commitment_loss,\n            psnr=psnr,\n            ssim=ssim,\n            codebook_usage=codebook_usage,\n        )\n        return loss, (outputs[""recon""], metrics)\n\n    @nnx.jit(donate_argnums=0)\n    def train_step(\n        optimizer: nnx.ModelAndOptimizer, inputs: dict\n    ) -> tuple[jax.Array, jax.Array, dict]:\n        def loss_fn(model: TokenizerVQVAE) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n            model.train()\n            return tokenizer_loss_fn(model, inputs, training=True)\n\n        (loss, (recon, metrics)), grads = nnx.value_and_grad(loss_fn, has_aux=True)(\n            optimizer.model\n        )\n        optimizer.update(grads)\n        if args.log_gradients:\n            metrics[""encoder_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""encoder""]\n            )\n            metrics[""vq_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""vq""]\n            )\n            metrics[""decoder_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""decoder""]\n            )\n        return loss, recon, metrics\n\n    @nnx.jit\n    def val_step(\n        tokenizer: TokenizerVQVAE, inputs: dict\n    ) -> tuple[jax.Array, jax.Array, dict]:\n        tokenizer.eval()\n        (loss, (recon, metrics)) = tokenizer_loss_fn(tokenizer, inputs, training=False)\n        return loss, recon, metrics\n\n    def calculate_validation_metrics(val_dataloader, tokenizer):\n        step = 0\n        loss_per_step = []\n        metrics_per_step = []\n        batch = None\n        recon = None\n        for batch in val_dataloader:\n            loss, recon, metrics = val_step(tokenizer, batch)\n            loss_per_step.append(loss)\n            metrics_per_step.append(metrics)\n            step += 1\n            if step > args.val_steps:\n                break\n\n        if step < args.val_steps:\n            print(\n                f""Warning: Your validation dataset is too small to make val_steps many steps. Made {step} steps, expected {args.val_steps}""\n            )\n\n        val_loss = np.mean(loss_per_step)\n        val_metrics = {\n            f""val_{key}"": np.mean([float(m[key]) for m in metrics_per_step])\n            for key in metrics_per_step[0].keys()\n        }\n        val_metrics[""val_loss""] = val_loss\n        return val_metrics, batch, recon\n\n    # --- TRAIN LOOP ---\n    dataloader_train = (\n        {\n            ""videos"": jax.make_array_from_process_local_data(\n                videos_sharding, elem[""videos""]\n            ),\n        }\n        for elem in train_iterator\n    )\n    dataloader_val = None\n    if val_iterator:\n        dataloader_val = (\n            {\n                ""videos"": jax.make_array_from_process_local_data(\n                    videos_sharding, elem[""videos""]\n                ),\n            }\n            for elem in val_iterator\n        )\n    if jax.process_index() == 0:\n        first_batch = next(dataloader_train)\n        compiled = train_step.lower(optimizer, first_batch).compile()\n        print_compiled_memory_stats(compiled.memory_analysis())\n        print_compiled_cost_analysis(compiled.cost_analysis())\n        # Do not skip the first batch during training\n        dataloader_train = itertools.chain([first_batch], dataloader_train)\n    print(f""Starting training from step {step}..."")\n    first_step = step\n    while step < args.num_steps:\n        for batch in dataloader_train:\n            # --- Train step ---\n            loss, recon, metrics = train_step(optimizer, batch)\n            if step == first_step:\n                print_mem_stats(""After params initialized"")\n            step += 1\n\n            # --- Validation loss ---\n            val_results = {}\n            if dataloader_val and step % args.val_interval == 0:\n                print(""Calculating validation metrics..."")\n                val_metrics, val_gt_batch, val_recon = calculate_validation_metrics(\n                    dataloader_val, optimizer.model\n                )\n                print(f""Step {step}, validation loss: {val_metrics['val_loss']}"")\n                val_results = {\n                    ""metrics"": val_metrics,\n                    ""gt_batch"": val_gt_batch,\n                    ""recon"": val_recon,\n                }\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    log_dict = {""loss"": loss, ""step"": step, **metrics}\n                    if val_results:\n                        log_dict.update(val_results[""metrics""])\n                    wandb.log(log_dict)\n                if step % args.log_image_interval == 0:\n                    gt_seq = batch[""videos""][0].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    if val_results and step % args.val_interval == 0:\n                        val_results[""gt_seq_val""] = (\n                            val_results[""gt_batch""][""videos""][0].astype(jnp.float32)\n                            / 255.0\n                        )\n                        val_results[""recon_seq_val""] = val_results[""recon""][0].clip(\n                            0, 1\n                        )\n                        val_results[""val_comparison_seq""] = jnp.concatenate(\n                            (val_results[""gt_seq_val""], val_results[""recon_seq_val""]),\n                            axis=1,\n                        )\n                        val_results[""val_comparison_seq""] = einops.rearrange(\n                            val_results[""val_comparison_seq""] * 255,\n                            ""t h w c -> h (t w) c"",\n                        )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[0])),\n                            recon=wandb.Image(np.asarray(recon_seq[0])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        if val_results and step % args.val_interval == 0:\n                            log_images.update(\n                                dict(\n                                    val_image=wandb.Image(\n                                        np.asarray(val_results[""gt_seq_val""][0])\n                                    ),\n                                    val_recon=wandb.Image(\n                                        np.asarray(val_results[""recon_seq_val""][0])\n                                    ),\n                                    val_true_vs_recon=wandb.Image(\n                                        np.asarray(\n                                            val_results[""val_comparison_seq""].astype(\n                                                np.uint8\n                                            )\n                                        )\n                                    ),\n                                )\n                            )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                assert checkpoint_manager is not None\n                optimizer_state = nnx.state(optimizer)\n                if val_iterator:\n                    ckpt_manager_args = ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        train_dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            train_iterator  # type: ignore\n                        ),\n                        val_dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            val_iterator  # type: ignore\n                        ),\n                    )\n                else:\n                    ckpt_manager_args = ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        train_dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            train_iterator  # type: ignore\n                        ),\n                    )\n                checkpoint_manager.save(step, args=ckpt_manager_args)\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    if checkpoint_manager:\n        checkpoint_manager.close()\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(Args)\n    main(args)\n",python,tab
+13,598913,"jasmine/train_tokenizer.py",5065,0,"",python,selection_command
+14,602099,"jasmine/train_tokenizer.py",5065,23,"        num_workers=8,\n",python,content
+15,614644,"jasmine/utils/dataloader.py",0,0,"",python,tab
+16,617906,"jasmine/train_tokenizer.py",0,0,"",python,tab

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-a05e2f67-e3d7-4b95-a2cd-0a3457dff2f71757085303333-2025_09_05-17.15.08.473/source.csv

@@ -0,0 +1,167 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",0,0,"#!/usr/bin/env bash\n\n#SBATCH --nodes=1\n#SBATCH --ntasks-per-node=1\n#SBATCH --time=24:00:00\n#SBATCH --cpus-per-task=8\n#SBATCH --gres=gpu:1\n#SBATCH --output=/fast/project/HFMI_SynergyUnit/jafar_ws/logs/franz/coinrun/tokenizer/%x_%j.log\n#SBATCH --error=/fast/project/HFMI_SynergyUnit/jafar_ws/logs/franz/coinrun/tokenizer/%x_%j.log\n#SBATCH --job-name=train_tokenizer_coinrun_reproduction_q-loss-fix\n\n# Log the sbatch script\ncat $0\n\nmodule unload mpi/openmpi/5.0\nmodule unload devel/cuda/12.4\nsource .venv/bin/activate\n\njob_name=$SLURM_JOB_NAME\nslurm_job_id=$SLURM_JOB_ID\n\ntags=""coinrun reproduction 10m_dataset tokenizer q-loss-fix""\n\narray_records_dir=""/fast/project/HFMI_SynergyUnit/jafar_ws/data/coinrun/array_records_10m""\n\nCHECKPOINT_DIR=""/fast/project/HFMI_SynergyUnit/jafar_ws/checkpoints/coinrun/tokenizer/${job_name}_${slurm_job_id}""\nmkdir -p $CHECKPOINT_DIR\n\nenv | grep SLURM\n\nsrun python train_tokenizer.py \\n    --wandb_id $SLURM_JOB_ID \\n    --ckpt_dir $CHECKPOINT_DIR \\n    --batch_size=48 \\n    --image_height=64 \\n    --image_width=64 \\n    --init_lr=0 \\n    --max_lr=3e-4 \\n    --log_image_interval=1000 \\n    --log_checkpoint_interval=1000 \\n    --log \\n    --name=""${job_name}_${slurm_job_id}"" \\n    --tags $tags \\n    --entity instant-uv \\n    --project jafar \\n    --data_dir $array_records_dir &\n\nchild_pid=$!\n\nwait $child_pid\n\n",shellscript,tab
+2,147,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"5:15:08 PM [info] Activating crowd-code\n5:15:08 PM [info] Recording started\n5:15:08 PM [info] Initializing git provider using file system watchers...\n5:15:08 PM [info] Git repository found\n5:15:08 PM [info] Git provider initialized successfully\n5:15:08 PM [info] Initial git state: [object Object]\n",Log,tab
+3,1172,"TERMINAL",0,0,"",,terminal_focus
+4,1173,"TERMINAL",0,0,"",,terminal_focus
+5,1308,"TERMINAL",0,0,"",,terminal_focus
+6,1309,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",0,0,"",shellscript,tab
+7,1309,"TERMINAL",0,0,"",,terminal_focus
+8,1583,"TERMINAL",0,0,"source /home/franz.srambical/jafar/.venv/bin/activate",,terminal_command
+9,5259,"TERMINAL",0,0,"squeue",,terminal_command
+10,5263,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  28749 franz.sram interacti     1     2  R 2025-09-05T15:07:50 2025-09-05T15:07:50  2:07:23 1-00:00:00 hai008\r\n  28086   xiao.liu interacti     1    64  R 2025-09-05T06:35:30 2025-09-05T06:35:30 10:39:43   23:59:00 hai001\r\n  28085   xiao.liu interacti     1    64  R 2025-09-05T01:22:59 2025-09-05T01:22:59 15:52:14   23:59:00 hai005\r\n  28084   xiao.liu interacti     1    64  R 2025-09-05T01:22:55 2025-09-05T01:22:55 15:52:18   23:59:00 hai002\r\n  28754 franz.sram  standard     1    16  R 2025-09-05T15:45:48 2025-09-05T15:45:48  1:29:25 1-00:00:00 hai003\r\n  28753 yoland.sav  standard     1   200  R 2025-09-05T15:15:33 2025-09-05T15:15:46  1:59:27    4:00:00 hai008\r\n  28747 alfred.ngu  standard     1    10  R 2025-09-05T14:11:49 2025-09-05T14:12:06  3:03:07 1-00:00:00 hai006\r\n  28748 alfred.ngu  standard     1    10  R 2025-09-05T14:11:59 2025-09-05T14:12:06  3:03:07 1-00:00:00 hai007\r\n  28746 alfred.ngu  standard     1    10  R 2025-09-05T14:11:45 2025-09-05T14:11:45  3:03:28 1-00:00:00 hai004\r\n  28591 alfred.ngu  standard     1    10  R 2025-09-05T12:41:52 2025-09-05T12:41:52  4:33:21 1-00:00:00 hai003\r\n]0;franz.srambical@hai-login1:~/jafar",,terminal_output
+11,10329,"TERMINAL",0,0,"git log",,terminal_command
+12,10378,"TERMINAL",0,0,"]633;C",,terminal_output
+13,10575,"TERMINAL",0,0,"[?1h=\rcommit 9b699287faf719af5fd81037717177b3825d32b3 (HEAD -> unnormalized-emb-in-q-loss, origin/unnormalized-emb-in-q-loss)\r\nAuthor: emergenz <franz.srambical@gmail.com>\r\nDate:   Thu Sep 4 20:13:42 2025 +0200\r\n\r\n    fix: q-loss calculation with unnormalized emb\r\n\r\ncommit bbef694ce01972ca6d625aa8e4d9e5c0d3f95305 (origin/main, origin/HEAD, main)\r\nAuthor: Franz Srambical <79149449+emergenz@users.noreply.github.com>\r\nDate:   Wed Aug 27 13:18:33 2025 +0200\r\n\r\n    chore: simplified parameter calculation (#174)\r\n    \r\n    * feat: print cost and memory analysis of train_step\r\n    \r\n    * chore: pre-commit\r\n    \r\n    * chore: only return grain_iterator\r\n    \r\n    * feat: donate optimizer buffer during nnx.jit\r\n:",,terminal_output
+14,14792,"TERMINAL",0,0,"\r[?1l>]0;franz.srambical@hai-login1:~/jafar",,terminal_output
+15,25094,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",0,0,"Switched from branch 'unnormalized-emb-in-q-loss' to 'main'",shellscript,git_branch_checkout
+16,30708,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",20,0,"",shellscript,selection_command
+17,30951,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",21,0,"",shellscript,selection_command
+18,30972,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",39,0,"",shellscript,selection_command
+19,31005,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",67,0,"",shellscript,selection_command
+20,31038,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",91,0,"",shellscript,selection_command
+21,31072,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",117,0,"",shellscript,selection_command
+22,31215,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",138,0,"",shellscript,selection_command
+23,31351,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",234,0,"",shellscript,selection_command
+24,31505,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",329,0,"",shellscript,selection_command
+25,32058,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",330,0,"",shellscript,selection_command
+26,32304,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",337,0,"",shellscript,selection_command
+27,32328,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",339,0,"",shellscript,selection_command
+28,32370,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",342,0,"",shellscript,selection_command
+29,32394,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",343,0,"",shellscript,selection_command
+30,32773,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",395,0,"",shellscript,selection_command
+31,35175,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",394,0,"",shellscript,selection_command
+32,35416,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",392,0,"",shellscript,selection_command
+33,35567,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",391,0,"",shellscript,selection_command
+34,35688,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",387,0,"",shellscript,selection_command
+35,35844,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",386,0,"",shellscript,selection_command
+36,36444,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",385,0,"",shellscript,selection_command
+37,36733,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",385,0,"n",shellscript,content
+38,36734,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",386,0,"",shellscript,selection_keyboard
+39,36798,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",386,0,"o",shellscript,content
+40,36798,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",387,0,"",shellscript,selection_keyboard
+41,36893,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",386,0,"",shellscript,selection_command
+42,37289,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",387,0,"",shellscript,selection_command
+43,37403,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",387,0,"-",shellscript,content
+44,37403,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",388,0,"",shellscript,selection_keyboard
+45,37524,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",387,0,"",shellscript,selection_command
+46,39733,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",399,0,"",shellscript,selection_command
+47,39985,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",422,0,"",shellscript,selection_command
+48,40006,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",429,0,"",shellscript,selection_command
+49,40038,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",431,0,"",shellscript,selection_command
+50,40072,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",460,0,"",shellscript,selection_command
+51,40104,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",490,0,"",shellscript,selection_command
+52,40138,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",516,0,"",shellscript,selection_command
+53,40171,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",518,0,"",shellscript,selection_command
+54,40205,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",542,0,"",shellscript,selection_command
+55,40237,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",569,0,"",shellscript,selection_command
+56,40274,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",571,0,"",shellscript,selection_command
+57,40482,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",630,0,"",shellscript,selection_command
+58,41334,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",628,0,"",shellscript,selection_command
+59,41485,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",627,0,"",shellscript,selection_command
+60,41616,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",623,0,"",shellscript,selection_command
+61,41759,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",622,0,"",shellscript,selection_command
+62,42879,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",621,0,"",shellscript,selection_command
+63,43219,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",621,0,"n",shellscript,content
+64,43220,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",622,0,"",shellscript,selection_keyboard
+65,43231,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",622,0,"o",shellscript,content
+66,43231,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",623,0,"",shellscript,selection_keyboard
+67,43446,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",623,0,"-",shellscript,content
+68,43447,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",624,0,"",shellscript,selection_keyboard
+69,43507,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",623,0,"",shellscript,selection_command
+70,44444,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",572,0,"",shellscript,selection_command
+71,46500,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",1220,0,"",shellscript,selection_command
+72,46684,"slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",1351,0,"",shellscript,selection_command
+73,48737,"TERMINAL",0,0,"sbatch slurm/jobs/franz/berlin/coinrun/coinrun_tokenizer/coinrun_tokenizer_repoduction.sbatch",,terminal_command
+74,48738,"TERMINAL",0,0,"]633;CSubmitted batch job 28755\r\n]0;franz.srambical@hai-login1:~/jafar",,terminal_output
+75,51579,"TERMINAL",0,0,"squeue",,terminal_command
+76,51582,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  28749 franz.sram interacti     1     2  R 2025-09-05T15:07:50 2025-09-05T15:07:50  2:08:09 1-00:00:00 hai008\r\n  28086   xiao.liu interacti     1    64  R 2025-09-05T06:35:30 2025-09-05T06:35:30 10:40:29   23:59:00 hai001\r\n  28085   xiao.liu interacti     1    64  R 2025-09-05T01:22:59 2025-09-05T01:22:59 15:53:00   23:59:00 hai005\r\n  28084   xiao.liu interacti     1    64  R 2025-09-05T01:22:55 2025-09-05T01:22:55 15:53:04   23:59:00 hai002\r\n  28755 franz.sram  standard     1    16  R 2025-09-05T17:15:57 2025-09-05T17:15:57     0:02 1-00:00:00 hai004\r\n  28754 franz.sram  standard     1    16  R 2025-09-05T15:45:48 2025-09-05T15:45:48  1:30:11 1-00:00:00 hai003\r\n  28753 yoland.sav  standard     1   200  R 2025-09-05T15:15:33 2025-09-05T15:15:46  2:00:13    4:00:00 hai008\r\n  28747 alfred.ngu  standard     1    10  R 2025-09-05T14:11:49 2025-09-05T14:12:06  3:03:53 1-00:00:00 hai006\r\n  28748 alfred.ngu  standard     1    10  R 2025-09-05T14:11:59 2025-09-05T14:12:06  3:03:53 1-00:00:00 hai007\r\n  28746 alfred.ngu  standard     1    10  R 2025-09-05T14:11:45 2025-09-05T14:11:45  3:04:14 1-00:00:00 hai004\r\n  28591 alfred.ngu  standard     1    10  R 2025-09-05T12:41:52 2025-09-05T12:41:52  4:34:07 1-00:00:00 hai003\r\n]0;franz.srambical@hai-login1:~/jafar",,terminal_output
+77,79382,"utils/nn.py",0,0,"import math\nfrom typing import Tuple, Callable, List\n\nfrom flax import nnx\nimport jax\nimport jax.numpy as jnp\nimport einops\n\n\nclass SpatioTemporalPositionalEncoding(nnx.Module):\n    """"""\n    Applies separate sinusoidal positional encodings to the temporal and spatial dimensions.\n    """"""\n\n    def __init__(self, d_model: int, max_len: int = 5000):\n        self.d_model = d_model\n        self.max_len = max_len\n\n        pe = jnp.zeros((self.max_len, self.d_model))\n        position = jnp.arange(0, self.max_len, dtype=jnp.float32)[:, None]\n        div_term = jnp.exp(\n            jnp.arange(0, self.d_model, 2) * (-math.log(10000.0) / self.d_model)\n        )\n        pe = pe.at[:, 0::2].set(jnp.sin(position * div_term))\n        pe = pe.at[:, 1::2].set(jnp.cos(position * div_term))\n        self.pe = nnx.Variable(pe)\n\n    def __call__(self, x: jax.Array) -> jax.Array:\n        """"""\n        Args:\n            x: The input tensor of shape (Batch, Time, Space, Dimension).\n\n        Returns:\n            The input tensor with positional encodings added.\n        """"""\n        assert x.ndim == 4, f""Input must be 4-dimensional, but got shape {x.shape}""\n\n        num_timesteps = x.shape[1]\n        num_spatial_patches = x.shape[2]\n\n        # Temporal positional encoding: (1, T, 1, D)\n        temporal_pe = self.pe.value[None, :num_timesteps, None, :]\n        x = x + temporal_pe\n\n        # Spatial positional encoding: (1, 1, S, D)\n        spatial_pe = self.pe.value[None, None, :num_spatial_patches, :]\n        x = x + spatial_pe\n\n        return x\n\n\nclass STBlock(nnx.Module):\n    def __init__(\n        self,\n        dim: int,\n        ffn_dim: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        rngs: nnx.Rngs,\n        sow_weights: bool,\n        sow_activations: bool,\n    ):\n        self.dim = dim\n        self.ffn_dim = ffn_dim\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.sow_weights = sow_weights\n        self.sow_activations = sow_activations\n\n        self.spatial_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n        self.spatial_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.dim,\n            qkv_features=self.dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=False\n            ),\n            rngs=rngs,\n            decode=False,\n        )\n\n        self.temporal_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n        self.temporal_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.dim,\n            qkv_features=self.dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=False,\n        )\n\n        self.ffn_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n        self.ffn_dense1 = nnx.Linear(\n            in_features=self.dim,\n            out_features=self.ffn_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense2 = nnx.Linear(\n            in_features=self.ffn_dim,\n            out_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    @nnx.remat\n    def __call__(self, x_BTNM: jax.Array) -> jax.Array:\n        # --- Spatial attention ---\n        z_BTNM = self.spatial_norm(x_BTNM)\n        z_BTNM = self.spatial_attention(z_BTNM, sow_weights=self.sow_weights)\n        x_BTNM = x_BTNM + z_BTNM\n\n        # --- Temporal attention ---\n        x_BNTM = x_BTNM.swapaxes(1, 2)\n        z_BNTM = self.temporal_norm(x_BNTM)\n        z_BNTM = self.temporal_attention(z_BNTM, sow_weights=self.sow_weights)\n        x_BNTM = x_BNTM + z_BNTM\n        x_BTNM = x_BNTM.swapaxes(1, 2)\n\n        # --- Feedforward ---\n        z_BTNM = self.ffn_norm(x_BTNM)\n        z_BTND = self.ffn_dense1(z_BTNM)\n        z_BTND = jax.nn.gelu(z_BTND)\n        z_BTNM = self.ffn_dense2(z_BTND)\n        x_BTNM = x_BTNM + z_BTNM\n        if self.sow_activations:\n            self.sow(nnx.Intermediate, ""activations"", x_BTNM)\n        return x_BTNM\n\n\nclass STTransformer(nnx.Module):\n    """"""\n    Dimension keys:\n        B: batch size\n        T: number of frames\n        N: number of patches per frame\n        I: number of input features\n        M: model dimension\n        D: FFN dimension\n        V: vocabulary size\n    """"""\n\n    def __init__(\n        self,\n        input_dim: int,\n        model_dim: int,\n        ffn_dim: int,\n        out_dim: int,\n        num_blocks: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        rngs: nnx.Rngs,\n        sow_weights: bool = False,\n        sow_activations: bool = False,\n        sow_logits: bool = False,\n        max_len: int = 5000,\n    ):\n        self.input_dim = input_dim\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.out_dim = out_dim\n        self.num_blocks = num_blocks\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.sow_logits = sow_logits\n        self.sow_weights = sow_weights\n        self.sow_activations = sow_activations\n\n        self.input_norm1 = nnx.LayerNorm(\n            num_features=self.input_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n        self.input_dense = nnx.Linear(\n            in_features=self.input_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_norm2 = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n\n        self.pos_enc = SpatioTemporalPositionalEncoding(self.model_dim, max_len=max_len)\n\n        self.blocks = []\n        for _ in range(self.num_blocks):\n            self.blocks.append(\n                STBlock(\n                    dim=self.model_dim,\n                    ffn_dim=self.ffn_dim,\n                    num_heads=self.num_heads,\n                    dropout=self.dropout,\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                    use_flash_attention=self.use_flash_attention,\n                    rngs=rngs,\n                    sow_weights=self.sow_weights,\n                    sow_activations=self.sow_activations,\n                )\n            )\n\n        self.output_dense = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.out_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    def __call__(self, x_BTNI: jax.Array) -> jax.Array:\n        x_BTNI = self.input_norm1(x_BTNI)\n        x_BTNM = self.input_dense(x_BTNI)\n        x_BTNM = self.input_norm2(x_BTNM)\n        x_BTNM = self.pos_enc(x_BTNM)\n        for block in self.blocks:\n            x_BTNM = block(x_BTNM)\n\n        x_BTNV = self.output_dense(x_BTNM)\n        if self.sow_logits:\n            self.sow(nnx.Intermediate, ""logits"", x_BTNV)\n        return x_BTNV\n\n\nclass TransformerBlock(nnx.Module):\n    def __init__(\n        self,\n        model_dim: int,\n        ffn_dim: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        decode: bool,\n        rngs: nnx.Rngs,\n        sow_weights: bool,\n        sow_activations: bool,\n    ):\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.decode = decode\n        self.sow_weights = sow_weights\n        self.sow_activations = sow_activations\n\n        self.temporal_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n        self.spatial_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n        self.ffn_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n        self.temporal_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.model_dim,\n            qkv_features=self.model_dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=self.decode,\n        )\n        self.spatial_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.model_dim,\n            qkv_features=self.model_dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=self.decode,\n        )\n        self.ffn_dense1 = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.ffn_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense2 = nnx.Linear(\n            in_features=self.ffn_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    @nnx.remat\n    def __call__(\n        self, x_BTNM: jax.Array, pos_index: Tuple[jax.Array, jax.Array] | None = None\n    ) -> jax.Array:\n        # --- Spatial attention ---\n        B, T, N, M = x_BTNM.shape\n        z_FNM = einops.rearrange(x_BTNM, ""b t n m -> (b t) n m"")\n        z_FNM = self.spatial_norm(z_FNM)\n        z_FNM = self.spatial_attention(z_FNM, sow_weights=self.sow_weights)\n        z_BTNM = einops.rearrange(z_FNM, ""(b t) n m -> b t n m"", t=T)\n        x_BTNM = x_BTNM + z_BTNM\n        # --- Temporal attention ---\n        z_PTM = einops.rearrange(x_BTNM, ""b t n m -> (b n) t m"")\n        z_PTM = self.temporal_norm(z_PTM)\n        z_PTM = self.temporal_attention(z_PTM, sow_weights=self.sow_weights)\n        z_BTNM = einops.rearrange(z_PTM, ""(b n) t m -> b t n m"", n=N)\n        x_BTNM = x_BTNM + z_BTNM\n        # --- Feedforward ---\n        z_BTNM = self.ffn_norm(x_BTNM)\n        z_BTND = self.ffn_dense1(z_BTNM)\n        z_BTND = jax.nn.gelu(z_BTND)\n        z_BTNM = self.ffn_dense2(z_BTND)\n        x_BTNM = x_BTNM + z_BTNM\n        if self.sow_activations:\n            self.sow(nnx.Intermediate, ""activations"", x_BTNM)\n\n        return x_BTNM\n\n\nclass Transformer(nnx.Module):\n    """"""\n    Dimension keys:\n        B: batch size\n        T: number of frames\n        N: number of patches per frame\n        I: number of input features\n        M: model dimension\n        D: FFN dimension\n        V: vocabulary size\n        F: number of frames in batch\n        P: number of patch positions in batch\n    """"""\n\n    def __init__(\n        self,\n        input_dim: int,\n        model_dim: int,\n        ffn_dim: int,\n        out_dim: int,\n        num_blocks: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        decode: bool,\n        rngs: nnx.Rngs,\n        sow_logits: bool = False,\n        sow_weights: bool = False,\n        sow_activations: bool = False,\n        max_len: int = 5000,\n    ):\n        self.input_dim = input_dim\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.out_dim = out_dim\n        self.num_blocks = num_blocks\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.sow_logits = sow_logits\n        self.sow_weights = sow_weights\n        self.sow_activations = sow_activations\n\n        self.input_norm1 = nnx.LayerNorm(\n            num_features=self.input_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n        self.input_dense = nnx.Linear(\n            in_features=self.input_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_norm2 = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n\n        self.pos_enc = SpatioTemporalPositionalEncoding(self.model_dim, max_len=max_len)\n\n        self.blocks: List[TransformerBlock] = []\n        for _ in range(self.num_blocks):\n            self.blocks.append(\n                TransformerBlock(\n                    model_dim=self.model_dim,\n                    ffn_dim=self.ffn_dim,\n                    num_heads=self.num_heads,\n                    dropout=self.dropout,\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                    use_flash_attention=self.use_flash_attention,\n                    decode=decode,\n                    sow_weights=self.sow_weights,\n                    sow_activations=self.sow_activations,\n                    rngs=rngs,\n                )\n            )\n        self.output_dense = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.out_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    def __call__(\n        self, x_BTNI: jax.Array, pos_index: Tuple[jax.Array, jax.Array] | None = None\n    ) -> jax.Array:\n        x_BTNI = self.input_norm1(x_BTNI)\n        x_BTNM = self.input_dense(x_BTNI)\n        x_BTNM = self.input_norm2(x_BTNM)\n        x_BTNM = self.pos_enc(x_BTNM)\n        for block in self.blocks:\n            x_BTNM = block(x_BTNM, pos_index)\n\n        x_BTNV = self.output_dense(x_BTNM)\n        if self.sow_logits:\n            self.sow(nnx.Intermediate, ""logits"", x_BTNV)\n        return x_BTNV\n\n\ndef normalize(x: jax.Array) -> jax.Array:\n    return x / (jnp.linalg.norm(x, ord=2, axis=-1, keepdims=True) + 1e-8)\n\n\nclass VectorQuantizer(nnx.Module):\n    """"""\n    Dimension keys:\n        D: B * T * N\n        K: number of latents\n        L: latent dimension\n    """"""\n\n    def __init__(\n        self,\n        latent_dim: int,\n        num_latents: int,\n        dropout: float,\n        dtype: jnp.dtype,\n        rngs: nnx.Rngs,\n    ):\n        self.latent_dim = latent_dim\n        self.num_latents = num_latents\n        self.dropout = dropout\n        self.dtype = dtype\n\n        self.codebook = nnx.Param(\n            normalize(\n                nnx.initializers.lecun_uniform()(\n                    rngs.params(), (self.num_latents, self.latent_dim)\n                )\n            )\n        )\n        self.drop = nnx.Dropout(self.dropout, rngs=rngs)\n\n    def __call__(\n        self, x_DL: jax.Array, training: bool\n    ) -> Tuple[jax.Array, jax.Array, jax.Array, jax.Array]:\n        # --- Compute distances ---\n        x_DL = x_DL.astype(self.dtype)\n        codebook = self.codebook.value.astype(self.dtype)\n\n        x_DL = normalize(x_DL)\n        normalized_codebook_KL = normalize(codebook)\n        distance_DK = -jnp.matmul(x_DL, normalized_codebook_KL.T)\n        if training:\n            distance_DK = self.drop(distance_DK)\n\n        # --- Get indices and embeddings ---\n        indices_D = jnp.argmin(distance_DK, axis=-1)\n        z_DL = codebook[indices_D]\n\n        # --- Straight through estimator ---\n        z_q_DL = x_DL + jax.lax.stop_gradient(z_DL - x_DL)\n        return z_q_DL, z_DL, x_DL, indices_D\n\n    def get_codes(self, indices_E: jax.Array) -> jax.Array:\n        return self.codebook[indices_E]\n\n\ndef _create_flash_attention_fn(use_flash_attention: bool, is_causal: bool) -> Callable:\n    """"""\n    Create an attention function that uses flash attention if enabled.\n\n    flax.nnx.MultiHeadAttention provides tensors with shape (batch..., length, num_heads, head_dim),\n    but jax.nn.dot_product_attention expects (batch, length, num_heads, head_dim). We reshape to\n    ensure compatibility. cuDNN's flash attention additionally requires a sequence length that\n    is a multiple of 4. We pad the sequence length to the nearest multiple of 4 and mask\n    accordingly. Note that cuDNN requires the mask to be broadcast before calling the attention\n    function due to strict shape checking.\n    """"""\n\n    def attention_fn(\n        query_BTHD, key_BSHD, value_BSHD, bias=None, mask_B111=None, **kwargs\n    ):\n        implementation = ""cudnn"" if use_flash_attention else None\n\n        def _merge_batch_dims(x):\n            return einops.rearrange(x, ""... l h k -> (...) l h k"")\n\n        def _pad(x, pad_size):\n            return jnp.pad(x, ((0, 0), (0, pad_size), (0, 0), (0, 0)))\n\n        original_shape = query_BTHD.shape\n        T = query_BTHD.shape[-3]\n        S = key_BSHD.shape[-3]\n\n        # Pad to nearest multiple of 4\n        Q = ((T + 3) // 4) * 4\n        pad_size_Q = Q - T\n        K = ((S + 3) // 4) * 4\n        pad_size_K = K - S\n\n        query_BQHD = _pad(_merge_batch_dims(query_BTHD), pad_size_Q)\n        key_BKHD = _pad(_merge_batch_dims(key_BSHD), pad_size_K)\n        value_BKHD = _pad(_merge_batch_dims(value_BSHD), pad_size_K)\n\n        attention_mask = jnp.ones((Q, K), dtype=jnp.bool_)\n        attention_mask = attention_mask.at[T:, :].set(False)\n        attention_mask = attention_mask.at[:, S:].set(False)\n\n        mask_11TS = attention_mask[jnp.newaxis, jnp.newaxis, :, :]\n\n        bias_4d = (\n            jnp.pad(\n                _merge_batch_dims(bias),\n                ((0, 0), (0, 0), (0, pad_size_Q), (0, pad_size_K)),\n            )\n            if bias is not None\n            else None\n        )\n\n        # NOTE: jax.nn.dot_product_attention does not support dropout\n        output_4d = jax.nn.dot_product_attention(\n            query=query_BQHD,\n            key=key_BKHD,\n            value=value_BKHD,\n            bias=bias_4d,\n            mask=mask_11TS,\n            implementation=implementation,\n            is_causal=is_causal,\n        )\n        return output_4d[..., :T, :, :].reshape(original_shape)\n\n    return attention_fn\n",python,tab
+78,80415,"utils/nn.py",16006,0,"",python,selection_command
+79,80665,"utils/nn.py",16016,0,"",python,selection_command
+80,80699,"utils/nn.py",16036,0,"",python,selection_command
+81,80729,"utils/nn.py",16057,0,"",python,selection_command
+82,80763,"utils/nn.py",16086,0,"",python,selection_command
+83,80796,"utils/nn.py",16112,0,"",python,selection_command
+84,80830,"utils/nn.py",16114,0,"",python,selection_command
+85,80863,"utils/nn.py",16123,0,"",python,selection_command
+86,80896,"utils/nn.py",16141,0,"",python,selection_command
+87,80930,"utils/nn.py",16155,0,"",python,selection_command
+88,80963,"utils/nn.py",16180,0,"",python,selection_command
+89,80996,"utils/nn.py",16206,0,"",python,selection_command
+90,81031,"utils/nn.py",16230,0,"",python,selection_command
+91,81064,"utils/nn.py",16256,0,"",python,selection_command
+92,81098,"utils/nn.py",16277,0,"",python,selection_command
+93,81131,"utils/nn.py",16287,0,"",python,selection_command
+94,81166,"utils/nn.py",16324,0,"",python,selection_command
+95,81200,"utils/nn.py",16363,0,"",python,selection_command
+96,81238,"utils/nn.py",16394,0,"",python,selection_command
+97,81267,"utils/nn.py",16413,0,"",python,selection_command
+98,81304,"utils/nn.py",16422,0,"",python,selection_command
+99,81334,"utils/nn.py",16457,0,"",python,selection_command
+100,81367,"utils/nn.py",16480,0,"",python,selection_command
+101,81400,"utils/nn.py",16530,0,"",python,selection_command
+102,81434,"utils/nn.py",16601,0,"",python,selection_command
+103,81468,"utils/nn.py",16619,0,"",python,selection_command
+104,81504,"utils/nn.py",16633,0,"",python,selection_command
+105,81535,"utils/nn.py",16643,0,"",python,selection_command
+106,81589,"utils/nn.py",16692,0,"",python,selection_command
+107,81688,"utils/nn.py",16643,0,"",python,selection_command
+108,81951,"utils/nn.py",16633,0,"",python,selection_command
+109,81974,"utils/nn.py",16619,0,"",python,selection_command
+110,82003,"utils/nn.py",16601,0,"",python,selection_command
+111,82038,"utils/nn.py",16530,0,"",python,selection_command
+112,82071,"utils/nn.py",16480,0,"",python,selection_command
+113,82105,"utils/nn.py",16457,0,"",python,selection_command
+114,82137,"utils/nn.py",16422,0,"",python,selection_command
+115,82172,"utils/nn.py",16413,0,"",python,selection_command
+116,82204,"utils/nn.py",16394,0,"",python,selection_command
+117,82237,"utils/nn.py",16363,0,"",python,selection_command
+118,82321,"utils/nn.py",16324,0,"",python,selection_command
+119,82322,"utils/nn.py",16287,0,"",python,selection_command
+120,82355,"utils/nn.py",16277,0,"",python,selection_command
+121,82372,"utils/nn.py",16256,0,"",python,selection_command
+122,82404,"utils/nn.py",16230,0,"",python,selection_command
+123,82437,"utils/nn.py",16206,0,"",python,selection_command
+124,82471,"utils/nn.py",16180,0,"",python,selection_command
+125,82504,"utils/nn.py",16155,0,"",python,selection_command
+126,82537,"utils/nn.py",16141,0,"",python,selection_command
+127,82570,"utils/nn.py",16123,0,"",python,selection_command
+128,82604,"utils/nn.py",16114,0,"",python,selection_command
+129,82637,"utils/nn.py",16112,0,"",python,selection_command
+130,82672,"utils/nn.py",16086,0,"",python,selection_command
+131,82704,"utils/nn.py",16057,0,"",python,selection_command
+132,82736,"utils/nn.py",16036,0,"",python,selection_command
+133,82771,"utils/nn.py",16016,0,"",python,selection_command
+134,82938,"utils/nn.py",16006,0,"",python,selection_command
+135,83081,"utils/nn.py",15973,0,"",python,selection_command
+136,83273,"utils/nn.py",15965,34,"class VectorQuantizer(nnx.Module):",python,selection_command
+137,83324,"utils/nn.py",15965,149,"class VectorQuantizer(nnx.Module):\n    """"""\n    Dimension keys:\n        D: B * T * N\n        K: number of latents\n        L: latent dimension\n    """"""\n",python,selection_command
+138,83482,"utils/nn.py",15965,448,"class VectorQuantizer(nnx.Module):\n    """"""\n    Dimension keys:\n        D: B * T * N\n        K: number of latents\n        L: latent dimension\n    """"""\n\n    def __init__(\n        self,\n        latent_dim: int,\n        num_latents: int,\n        dropout: float,\n        dtype: jnp.dtype,\n        rngs: nnx.Rngs,\n    ):\n        self.latent_dim = latent_dim\n        self.num_latents = num_latents\n        self.dropout = dropout\n        self.dtype = dtype\n",python,selection_command
+139,83655,"utils/nn.py",15965,727,"class VectorQuantizer(nnx.Module):\n    """"""\n    Dimension keys:\n        D: B * T * N\n        K: number of latents\n        L: latent dimension\n    """"""\n\n    def __init__(\n        self,\n        latent_dim: int,\n        num_latents: int,\n        dropout: float,\n        dtype: jnp.dtype,\n        rngs: nnx.Rngs,\n    ):\n        self.latent_dim = latent_dim\n        self.num_latents = num_latents\n        self.dropout = dropout\n        self.dtype = dtype\n\n        self.codebook = nnx.Param(\n            normalize(\n                nnx.initializers.lecun_uniform()(\n                    rngs.params(), (self.num_latents, self.latent_dim)\n                )\n            )\n        )\n        self.drop = nnx.Dropout(self.dropout, rngs=rngs)\n",python,selection_command
+140,83783,"utils/nn.py",15965,985,"class VectorQuantizer(nnx.Module):\n    """"""\n    Dimension keys:\n        D: B * T * N\n        K: number of latents\n        L: latent dimension\n    """"""\n\n    def __init__(\n        self,\n        latent_dim: int,\n        num_latents: int,\n        dropout: float,\n        dtype: jnp.dtype,\n        rngs: nnx.Rngs,\n    ):\n        self.latent_dim = latent_dim\n        self.num_latents = num_latents\n        self.dropout = dropout\n        self.dtype = dtype\n\n        self.codebook = nnx.Param(\n            normalize(\n                nnx.initializers.lecun_uniform()(\n                    rngs.params(), (self.num_latents, self.latent_dim)\n                )\n            )\n        )\n        self.drop = nnx.Dropout(self.dropout, rngs=rngs)\n\n    def __call__(\n        self, x_DL: jax.Array, training: bool\n    ) -> Tuple[jax.Array, jax.Array, jax.Array, jax.Array]:\n        # --- Compute distances ---\n        x_DL = x_DL.astype(self.dtype)\n        codebook = self.codebook.value.astype(self.dtype)\n",python,selection_command
+141,83922,"utils/nn.py",15965,1206,"class VectorQuantizer(nnx.Module):\n    """"""\n    Dimension keys:\n        D: B * T * N\n        K: number of latents\n        L: latent dimension\n    """"""\n\n    def __init__(\n        self,\n        latent_dim: int,\n        num_latents: int,\n        dropout: float,\n        dtype: jnp.dtype,\n        rngs: nnx.Rngs,\n    ):\n        self.latent_dim = latent_dim\n        self.num_latents = num_latents\n        self.dropout = dropout\n        self.dtype = dtype\n\n        self.codebook = nnx.Param(\n            normalize(\n                nnx.initializers.lecun_uniform()(\n                    rngs.params(), (self.num_latents, self.latent_dim)\n                )\n            )\n        )\n        self.drop = nnx.Dropout(self.dropout, rngs=rngs)\n\n    def __call__(\n        self, x_DL: jax.Array, training: bool\n    ) -> Tuple[jax.Array, jax.Array, jax.Array, jax.Array]:\n        # --- Compute distances ---\n        x_DL = x_DL.astype(self.dtype)\n        codebook = self.codebook.value.astype(self.dtype)\n\n        x_DL = normalize(x_DL)\n        normalized_codebook_KL = normalize(codebook)\n        distance_DK = -jnp.matmul(x_DL, normalized_codebook_KL.T)\n        if training:\n            distance_DK = self.drop(distance_DK)\n",python,selection_command
+142,84248,"utils/nn.py",15965,1340,"class VectorQuantizer(nnx.Module):\n    """"""\n    Dimension keys:\n        D: B * T * N\n        K: number of latents\n        L: latent dimension\n    """"""\n\n    def __init__(\n        self,\n        latent_dim: int,\n        num_latents: int,\n        dropout: float,\n        dtype: jnp.dtype,\n        rngs: nnx.Rngs,\n    ):\n        self.latent_dim = latent_dim\n        self.num_latents = num_latents\n        self.dropout = dropout\n        self.dtype = dtype\n\n        self.codebook = nnx.Param(\n            normalize(\n                nnx.initializers.lecun_uniform()(\n                    rngs.params(), (self.num_latents, self.latent_dim)\n                )\n            )\n        )\n        self.drop = nnx.Dropout(self.dropout, rngs=rngs)\n\n    def __call__(\n        self, x_DL: jax.Array, training: bool\n    ) -> Tuple[jax.Array, jax.Array, jax.Array, jax.Array]:\n        # --- Compute distances ---\n        x_DL = x_DL.astype(self.dtype)\n        codebook = self.codebook.value.astype(self.dtype)\n\n        x_DL = normalize(x_DL)\n        normalized_codebook_KL = normalize(codebook)\n        distance_DK = -jnp.matmul(x_DL, normalized_codebook_KL.T)\n        if training:\n            distance_DK = self.drop(distance_DK)\n\n        # --- Get indices and embeddings ---\n        indices_D = jnp.argmin(distance_DK, axis=-1)\n        z_DL = codebook[indices_D]\n",python,selection_command
+143,84419,"utils/nn.py",15965,1490,"class VectorQuantizer(nnx.Module):\n    """"""\n    Dimension keys:\n        D: B * T * N\n        K: number of latents\n        L: latent dimension\n    """"""\n\n    def __init__(\n        self,\n        latent_dim: int,\n        num_latents: int,\n        dropout: float,\n        dtype: jnp.dtype,\n        rngs: nnx.Rngs,\n    ):\n        self.latent_dim = latent_dim\n        self.num_latents = num_latents\n        self.dropout = dropout\n        self.dtype = dtype\n\n        self.codebook = nnx.Param(\n            normalize(\n                nnx.initializers.lecun_uniform()(\n                    rngs.params(), (self.num_latents, self.latent_dim)\n                )\n            )\n        )\n        self.drop = nnx.Dropout(self.dropout, rngs=rngs)\n\n    def __call__(\n        self, x_DL: jax.Array, training: bool\n    ) -> Tuple[jax.Array, jax.Array, jax.Array, jax.Array]:\n        # --- Compute distances ---\n        x_DL = x_DL.astype(self.dtype)\n        codebook = self.codebook.value.astype(self.dtype)\n\n        x_DL = normalize(x_DL)\n        normalized_codebook_KL = normalize(codebook)\n        distance_DK = -jnp.matmul(x_DL, normalized_codebook_KL.T)\n        if training:\n            distance_DK = self.drop(distance_DK)\n\n        # --- Get indices and embeddings ---\n        indices_D = jnp.argmin(distance_DK, axis=-1)\n        z_DL = codebook[indices_D]\n\n        # --- Straight through estimator ---\n        z_q_DL = x_DL + jax.lax.stop_gradient(z_DL - x_DL)\n        return z_q_DL, z_DL, x_DL, indices_D\n",python,selection_command
+144,84579,"utils/nn.py",15965,1591,"class VectorQuantizer(nnx.Module):\n    """"""\n    Dimension keys:\n        D: B * T * N\n        K: number of latents\n        L: latent dimension\n    """"""\n\n    def __init__(\n        self,\n        latent_dim: int,\n        num_latents: int,\n        dropout: float,\n        dtype: jnp.dtype,\n        rngs: nnx.Rngs,\n    ):\n        self.latent_dim = latent_dim\n        self.num_latents = num_latents\n        self.dropout = dropout\n        self.dtype = dtype\n\n        self.codebook = nnx.Param(\n            normalize(\n                nnx.initializers.lecun_uniform()(\n                    rngs.params(), (self.num_latents, self.latent_dim)\n                )\n            )\n        )\n        self.drop = nnx.Dropout(self.dropout, rngs=rngs)\n\n    def __call__(\n        self, x_DL: jax.Array, training: bool\n    ) -> Tuple[jax.Array, jax.Array, jax.Array, jax.Array]:\n        # --- Compute distances ---\n        x_DL = x_DL.astype(self.dtype)\n        codebook = self.codebook.value.astype(self.dtype)\n\n        x_DL = normalize(x_DL)\n        normalized_codebook_KL = normalize(codebook)\n        distance_DK = -jnp.matmul(x_DL, normalized_codebook_KL.T)\n        if training:\n            distance_DK = self.drop(distance_DK)\n\n        # --- Get indices and embeddings ---\n        indices_D = jnp.argmin(distance_DK, axis=-1)\n        z_DL = codebook[indices_D]\n\n        # --- Straight through estimator ---\n        z_q_DL = x_DL + jax.lax.stop_gradient(z_DL - x_DL)\n        return z_q_DL, z_DL, x_DL, indices_D\n\n    def get_codes(self, indices_E: jax.Array) -> jax.Array:\n        return self.codebook[indices_E]\n",python,selection_command
+145,85382,"utils/nn.py",17556,0,"",python,selection_command
+146,771760,"train_tokenizer.py",0,0,"import os\n\nos.environ.setdefault(""XLA_PYTHON_CLIENT_MEM_FRACTION"", ""0.98"")\n\nfrom dataclasses import dataclass, field\nfrom typing import cast, Optional\n\nimport einops\nimport itertools\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom models.tokenizer import TokenizerVQVAE\nfrom utils.dataloader import get_dataloader\nfrom utils.train_utils import (\n    get_lr_schedule,\n    count_parameters_by_component,\n    print_mem_stats,\n    print_compiled_memory_stats,\n    print_compiled_cost_analysis,\n)\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 300_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    vq_beta: float = 0.25\n    batch_size: int = 48\n    init_lr: float = 0.0\n    max_lr: float = 3e-4\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        20000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    warmup_steps: int = 10000\n    # Tokenizer\n    model_dim: int = 512\n    ffn_dim: int = 2048\n    latent_dim: int = 32\n    num_latents: int = 1024\n    patch_size: int = 4\n    num_blocks: int = 4\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.01\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_tokenizer""\n    tags: list[str] = field(default_factory=lambda: [""tokenizer""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n    log_checkpoint_keep_period: int = 20000\n    log_gradients: bool = False\n    wandb_id: str = """"\n    use_flash_attention: bool = True\n\n\ndef build_model(args: Args, rng: jax.Array) -> tuple[TokenizerVQVAE, jax.Array]:\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    return (\n        TokenizerVQVAE(\n            in_dim=args.image_channels,\n            model_dim=args.model_dim,\n            ffn_dim=args.ffn_dim,\n            latent_dim=args.latent_dim,\n            num_latents=args.num_latents,\n            patch_size=args.patch_size,\n            num_blocks=args.num_blocks,\n            num_heads=args.num_heads,\n            dropout=args.dropout,\n            codebook_dropout=args.codebook_dropout,\n            param_dtype=args.param_dtype,\n            dtype=args.dtype,\n            use_flash_attention=args.use_flash_attention,\n            rngs=rngs,\n        ),\n        rng,\n    )\n\n\ndef build_optimizer(\n    model: TokenizerVQVAE, args: Args\n) -> tuple[nnx.Optimizer, optax.Schedule]:\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.param_dtype,  # moments in full precision\n    )\n    optimizer = nnx.Optimizer(model, tx)\n    return optimizer, lr_schedule\n\n\ndef build_mesh_and_sharding(\n    num_devices: int,\n) -> tuple[Mesh, NamedSharding, NamedSharding]:\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n    return mesh, replicated_sharding, videos_sharding\n\n\ndef shard_optimizer_states(\n    optimizer: nnx.Optimizer, replicated_sharding: NamedSharding\n) -> None:\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n\ndef build_dataloader(args: Args) -> grain.DataLoaderIterator:\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n    return grain_iterator\n\n\ndef build_checkpoint_manager(args: Args) -> ocp.CheckpointManager:\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n    return checkpoint_manager\n\n\ndef restore_checkpoint_if_needed(\n    args: Args,\n    checkpoint_manager: ocp.CheckpointManager,\n    optimizer: nnx.Optimizer,\n    grain_iterator: grain.DataLoaderIterator,\n    restore_step: Optional[int] = None,\n) -> tuple[int, nnx.Optimizer, grain.DataLoaderIterator]:\n    step = 0\n    if restore_step is None:\n        restore_step = checkpoint_manager.latest_step()\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        restored = checkpoint_manager.restore(\n            restore_step,\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),  # type: ignore\n            ),\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        grain_iterator = restored[""dataloader_state""]\n        step = restore_step or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    return step, optimizer, grain_iterator\n\n\ndef main(args: Args) -> None:\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    tokenizer, rng = build_model(args, rng)\n\n    _, params, _ = nnx.split(tokenizer, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    optimizer, lr_schedule = build_optimizer(tokenizer, args)\n    del tokenizer\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    mesh, replicated_sharding, videos_sharding = build_mesh_and_sharding(num_devices)\n\n    shard_optimizer_states(optimizer, replicated_sharding)\n\n    # --- Initialize checkpoint manager ---\n    checkpoint_manager = build_checkpoint_manager(args)\n\n    # --- Create DataLoaderIterator from dataloader ---\n    grain_iterator = build_dataloader(args)\n\n    # --- Restore checkpoint ---\n    step, optimizer, grain_iterator = restore_checkpoint_if_needed(\n        args, checkpoint_manager, optimizer, grain_iterator\n    )\n\n    # --- Define loss and train step (close over args) ---\n    def tokenizer_loss_fn(\n        model: TokenizerVQVAE, inputs: dict\n    ) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n        gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n        inputs[""videos""] = gt.astype(args.dtype)\n        model.train()\n        outputs = model(inputs, training=True)\n        outputs[""recon""] = outputs[""recon""].astype(jnp.float32)\n        mse = jnp.square(gt - outputs[""recon""]).mean()\n        q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n        commitment_loss = jnp.square(\n            outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n        ).mean()\n        loss = mse + q_loss + args.vq_beta * commitment_loss\n\n        gt_clipped = gt.clip(0, 1).reshape(-1, *gt.shape[2:])\n        recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n        psnr = jnp.asarray(pix.psnr(gt_clipped, recon)).mean()\n        ssim = jnp.asarray(pix.ssim(gt_clipped, recon)).mean()\n        _, index_counts = jnp.unique_counts(\n            jnp.ravel(outputs[""indices""]), size=args.num_latents, fill_value=0\n        )\n        codebook_usage = (index_counts != 0).mean()\n        metrics = dict(\n            loss=loss,\n            mse=mse,\n            q_loss=q_loss,\n            commitment_loss=commitment_loss,\n            psnr=psnr,\n            ssim=ssim,\n            codebook_usage=codebook_usage,\n        )\n        return loss, (outputs[""recon""], metrics)\n\n    @nnx.jit(donate_argnums=0)\n    def train_step(\n        optimizer: nnx.Optimizer, inputs: dict\n    ) -> tuple[jax.Array, jax.Array, dict]:\n        def loss_fn(model: TokenizerVQVAE) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n            return tokenizer_loss_fn(model, inputs)\n\n        (loss, (recon, metrics)), grads = nnx.value_and_grad(loss_fn, has_aux=True)(\n            optimizer.model\n        )\n        optimizer.update(grads)\n        if args.log_gradients:\n            metrics[""encoder_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""encoder""]\n            )\n            metrics[""vq_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""vq""]\n            )\n            metrics[""decoder_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""decoder""]\n            )\n        return loss, recon, metrics\n\n    # --- TRAIN LOOP ---\n    dataloader = (\n        jax.make_array_from_process_local_data(videos_sharding, elem)\n        for elem in grain_iterator\n    )\n    if jax.process_index() == 0:\n        first_videos = next(dataloader)\n        sample_inputs = dict(videos=first_videos)\n        compiled = train_step.lower(optimizer, sample_inputs).compile()\n        print_compiled_memory_stats(compiled.memory_analysis())\n        print_compiled_cost_analysis(compiled.cost_analysis())\n        # Do not skip the first batch during training\n        dataloader = itertools.chain([first_videos], dataloader)\n    print(f""Starting training from step {step}..."")\n    first_step = step\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            inputs = dict(videos=videos)\n            loss, recon, metrics = train_step(optimizer, inputs)\n            if step == first_step:\n                print_mem_stats(""After params initialized"")\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[0])),\n                            recon=wandb.Image(np.asarray(recon_seq[0])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            grain_iterator  # type: ignore\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(Args)\n    main(args)\n",python,tab
+147,965095,"train_tokenizer.py",11371,0,"",python,selection_command
+148,965368,"train_tokenizer.py",9681,0,"",python,selection_command
+149,991765,"train_tokenizer.py",9708,0,"",python,selection_command
+150,991932,"train_tokenizer.py",9752,0,"",python,selection_command
+151,992098,"train_tokenizer.py",9708,0,"",python,selection_command
+152,992257,"train_tokenizer.py",9681,0,"",python,selection_command
+153,1000628,"train_tokenizer.py",9708,0,"",python,selection_command
+154,1000878,"train_tokenizer.py",9752,0,"",python,selection_command
+155,1000896,"train_tokenizer.py",9803,0,"",python,selection_command
+156,1000928,"train_tokenizer.py",9873,0,"",python,selection_command
+157,1000959,"train_tokenizer.py",9922,0,"",python,selection_command
+158,1000994,"train_tokenizer.py",9944,0,"",python,selection_command
+159,1001026,"train_tokenizer.py",9991,0,"",python,selection_command
+160,1001061,"train_tokenizer.py",10055,0,"",python,selection_command
+161,1001203,"train_tokenizer.py",10110,0,"",python,selection_command
+162,1148528,"train_tokenizer.py",10199,0,"",python,selection_command
+163,1148854,"train_tokenizer.py",10110,0,"",python,selection_command
+164,1187973,"train_tokenizer.py",10346,0,"",python,selection_command
+165,1188309,"train_tokenizer.py",1047,0,"",python,selection_command
+166,1189448,"train_tokenizer.py",10346,0,"",python,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-b22cb5a3-8f10-4e6e-9f4f-731b2b6708cc1764260897825-2025_11_27-17.28.25.85/source.csv

@@ -0,0 +1,45 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,4,"crowd-pilot/crowd_pilot/serialization_utils.py",0,0,"#!/usr/bin/env python3\n""""""\nCommon utilities for dataset serialization scripts.\n""""""\n\nfrom __future__ import annotations\n\nfrom dataclasses import dataclass, field\nfrom pathlib import Path\nfrom typing import List, Optional, Tuple, Dict, Any\n\nimport difflib\nimport re\nimport pandas as pd\nfrom datasets import Dataset, load_dataset\n\n\n_ANSI_CSI_RE = re.compile(r""\x1b\[[0-9;?]*[ -/]*[@-~]"")\n_ANSI_OSC_TERMINATED_RE = re.compile(r""\x1b\][\s\S]*?(?:\x07|\x1b\\)"")\n_ANSI_OSC_LINE_FALLBACK_RE = re.compile(r""\x1b\][^\n]*$"")\n_BRACKETED_PASTE_ENABLE = ""\x1b[?2004h""\n_BRACKETED_PASTE_DISABLE = ""\x1b[?2004l""\n_OSC_633 = ""\x1b]633;""\n_OSC_0 = ""\x1b]0;""\n\n\n\n@dataclass\nclass ConversationState:\n    """"""\n    Mutable state used while constructing conversations.\n    """"""\n    conversations: List[List[Dict[str, str]]]\n    max_tokens_per_conversation: int\n    max_tokens_per_message: int\n    min_conversation_messages: int\n    tokenizer: Any\n    conversation_token_counts: List[int] = field(default_factory=list)\n    current_conversation: List[Dict[str, str]] = field(default_factory=list)\n    current_tokens: int = 0\n    files_opened_in_conversation: set[str] = field(default_factory=set)\n\n    def finalize_conversation(self) -> None:\n        """"""\n        Finalize the current conversation: check constraints and append if valid.\n        Then reset state for the next conversation.\n        """"""\n        if self.current_conversation:\n            is_long_enough = len(self.current_conversation) >= self.min_conversation_messages\n            has_user = any(msg.get(""from"") == ""User"" for msg in self.current_conversation)\n            has_assistant = any(msg.get(""from"") == ""Assistant"" for msg in self.current_conversation)\n\n            if is_long_enough and has_user and has_assistant:\n                self.conversations.append(self.current_conversation)\n                self.conversation_token_counts.append(self.current_tokens)\n        \n        self.current_conversation = []\n        self.current_tokens = 0\n        self.files_opened_in_conversation.clear()\n\n    def append_message(self, message: Dict[str, str]) -> None:\n        value = message[""value""]\n        \n        tokens = self.tokenizer.encode(value)\n        num_tokens = len(tokens)\n\n        if num_tokens > self.max_tokens_per_message:\n            tokens = tokens[:self.max_tokens_per_message]\n            value = self.tokenizer.decode(tokens)\n            message[""value""] = value\n            num_tokens = self.max_tokens_per_message\n\n        if self.current_tokens + num_tokens > self.max_tokens_per_conversation:\n            self.finalize_conversation()\n\n        self.current_conversation.append(message)\n        self.current_tokens += num_tokens\n\n    def maybe_capture_file_contents(\n        self,\n        file_path: str,\n        content: str,\n    ) -> None:\n        """"""\n        Capture the contents of the given file in the current conversation if it hasn't been opened yet.\n        """"""\n        if file_path in self.files_opened_in_conversation:\n            return\n        cmd = f""cat -n {file_path}""\n        self.append_message({\n            ""from"": ""Assistant"",\n            ""value"": _fenced_block(""bash"", _clean_text(cmd)),\n        })\n        output = _line_numbered_output(content)\n        self.append_message({\n            ""from"": ""User"",\n            ""value"": f""<stdout>\n{output}\n</stdout>"",\n        })\n        self.files_opened_in_conversation.add(file_path)\n\n\ndef _clean_text(text: str) -> str:\n    # Normalize line endings and strip trailing spaces; preserve tabs/newlines.\n    return text.replace(""\r\n"", ""\n"").replace(""\r"", ""\n"").rstrip()\n\n\ndef _fenced_block(language: Optional[str], content: str) -> str:\n    lang = (language or """").lower()\n    return f""```{lang}\n{content}\n```\n""\n\n\ndef _apply_change(content: str, offset: int, length: int, new_text: str) -> str:\n    # Mirrors crowd_code_player.replay_file.apply_change\n    base = str(content)\n    text = str(new_text) if pd.notna(new_text) else """"\n    text = text.replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n    if offset > len(base):\n        base = base + ("" "" * (offset - len(base)))\n    return base[:offset] + text + base[offset + length:]\n\n\ndef _apply_backspaces(text: str) -> str:\n    out: List[str] = []\n    for ch in text:\n        if ch == ""\b"":  # \x08\n            if out:\n                out.pop()\n        else:\n            out.append(ch)\n    return """".join(out)\n\n\ndef _normalize_terminal_output(raw: str) -> str:\n    """"""\n    Normalize PTY/terminal output for training:\n      - Apply backspaces (\x08)\n      - Strip OSC (window title/shell integration) first, keeping BEL/ST terminators intact\n      - Resolve carriage returns (\r) by keeping the last rewrite per line\n      - Strip CSI (coloring etc.)\n      - Finally drop any remaining BEL (\x07)\n    """"""\n    if not raw:\n        return raw\n    s = _apply_backspaces(raw)\n    # Remove OSC sequences that are properly terminated (BEL or ST)\n    s = _ANSI_OSC_TERMINATED_RE.sub("""", s)\n    # Fallback: drop any unterminated OSC up to end-of-line only\n    s = ""\n"".join(_ANSI_OSC_LINE_FALLBACK_RE.sub("""", line) for line in s.split(""\n""))\n    # Resolve carriage returns per line:\n    # - If there are multiple rewrites, keep the last non-empty conversation\n    # - If it's CRLF (ending with '\r' before '\n'), keep the content before '\r'\n    resolved_lines: List[str] = []\n    for seg in s.split(""\n""):\n        parts = seg.split(""\r"")\n        chosen = """"\n        # pick last non-empty part if available; else last part\n        for p in reversed(parts):\n            if p != """":\n                chosen = p\n                break\n        if chosen == """" and parts:\n            chosen = parts[-1]\n        resolved_lines.append(chosen)\n    s = ""\n"".join(resolved_lines)\n    # Strip ANSI escape sequences\n    s = _ANSI_CSI_RE.sub("""", s)\n    # Remove any remaining BEL beeps\n    s = s.replace(""\x07"", """")\n    return s\n\n\ndef _line_numbered_output(content: str, start_line: Optional[int] = None, end_line: Optional[int] = None) -> str:\n    # FIXME (f.srambical): check whether this corresponds **exactly** to the output of cat -n {file_path} | sed -n '{vstart},{vend}p'\n    lines = content.splitlines()\n    total = len(lines)\n    if total == 0:\n        return """"\n    s = 1 if start_line is None else max(1, min(start_line, total))\n    e = total if end_line is None else max(1, min(end_line, total))\n    assert e >= s, ""End line number cannot be less than start line number! Likely a bug in the line numbering computation.""\n    buf: List[str] = []\n    for idx in range(s, e + 1):\n        buf.append(f""{idx:6}\t{lines[idx - 1]}"")\n    return ""\n"".join(buf)\n\n\ndef _compute_viewport(total_lines: int, center_line: int, radius: int) -> Tuple[int, int]:\n    if total_lines <= 0:\n        return (1, 0)\n    start = max(1, center_line - radius)\n    end = min(total_lines, center_line + radius)\n    assert end >= start, ""Viewport cannot have negative width! Likely a bug in the viewport computation.""\n    return (start, end)\n\n\ndef _escape_single_quotes_for_sed(text: str) -> str:\n    # Close quote, add an escaped single quote, reopen quote: '""'""'\n    return text.replace(""'"", ""'\""'\""'"")\n\n\ndef _compute_changed_block_lines(\n    before: str, after: str\n) -> Tuple[int, int, int, int, List[str]]:\n    """"""\n    Return 1-based start and end line numbers in 'before' that should be\n    replaced, 1-based start and end line numbers in 'after' that contain\n    the replacement, and the replacement lines from 'after'.\n\n    For pure deletions, the replacement list may be empty.\n    """"""\n    before_lines = before.splitlines()\n    after_lines = after.splitlines()\n    sm = difflib.SequenceMatcher(a=before_lines, b=after_lines, autojunk=False)\n    opcodes = [op for op in sm.get_opcodes() if op[0] != ""equal""]\n    assert opcodes, ""Opcode list cannot be empty! Likely a bug in the diff computation.""\n\n    first = opcodes[0]\n    last = opcodes[-1]\n    # i1/i2 refer to 'before' indices, j1/j2 to 'after'\n    start_before = max(1, first[1] + 1)\n    end_before = last[2]  # no increment since we go from 'exclusive' to 'inclusive' indexing\n    start_after = max(1, first[3] + 1)\n    end_after = last[4]\n    replacement_lines = after_lines[first[3] : last[4]]\n    return (start_before, end_before, start_after, end_after, replacement_lines)\n\n\ndef session_to_nemo_conversations(\n    df: pd.DataFrame,\n    max_tokens_per_conversation: int,\n    max_tokens_per_message: int,\n    min_conversation_messages: int,\n    tokenizer: Any,\n    viewport_radius: int = 10,\n    normalize_terminal_output: bool = True,\n    coalesce_radius: int = 5,\n) -> Tuple[List[List[Dict[str, str]]], List[int]]:\n    """"""\n    Convert a session DataFrame to one or more NeMo conversations.\n\n    - Conversations are created by approximately limiting the total tokens\n      across all `value` fields to `max_tokens_per_conversation`.\n    - When a new conversation starts (after the first), the first time a file is\n      referenced in that conversation we re-log the full file contents with\n      `cat -n <file>` and numbered output so that each conversation is self-contained.\n    """"""\n    file_states: Dict[str, str] = {}\n    per_file_viewport: Dict[str, Optional[Tuple[int, int]]] = {}\n\n    conversations: List[List[Dict[str, str]]] = []\n    conversation_token_counts: List[int] = []\n    conversation_state = ConversationState(\n        conversations=conversations,\n        conversation_token_counts=conversation_token_counts,\n        max_tokens_per_conversation=max_tokens_per_conversation,\n        max_tokens_per_message=max_tokens_per_message,\n        min_conversation_messages=min_conversation_messages,\n        tokenizer=tokenizer,\n    )\n\n    terminal_output_buffer: List[str] = []\n    pending_edits_before: Dict[str, Optional[str]] = {}\n    pending_edit_regions: Dict[str, Optional[Tuple[int, int]]] = {}\n\n    def _flush_terminal_output_buffer() -> None:\n        if not terminal_output_buffer:\n            return\n        aggregated = """".join(terminal_output_buffer)\n        out = aggregated\n        if normalize_terminal_output:\n            out = _normalize_terminal_output(out)\n        cleaned = _clean_text(out)\n        if cleaned.strip():\n            conversation_state.append_message({\n                ""from"": ""User"",\n                ""value"": f""<stdout>\n{cleaned}\n</stdout>"",\n            })\n        terminal_output_buffer.clear()\n\n    def _flush_pending_edit_for_file(target_file: str) -> None:\n        before_snapshot = pending_edits_before.get(target_file)\n        if before_snapshot is None:\n            return\n        after_state = file_states.get(target_file, """")\n        if before_snapshot.rstrip(""\n"") == after_state.rstrip(""\n""):\n            pending_edits_before[target_file] = None\n            pending_edit_regions[target_file] = None\n            return\n        (\n            start_before,\n            end_before,\n            start_after,\n            end_after,\n            repl_lines,\n        ) = _compute_changed_block_lines(before_snapshot, after_state)\n        before_total_lines = len(before_snapshot.splitlines())\n        if end_before < start_before:\n            escaped_lines = [_escape_single_quotes_for_sed(line) for line in repl_lines]\n            sed_payload = ""\n"".join(escaped_lines)\n            if start_before <= max(1, before_total_lines):\n                sed_cmd = f""sed -i '{start_before}i\\\n{sed_payload}' {target_file}""\n            else:\n                sed_cmd = f""sed -i '$a\\\n{sed_payload}' {target_file}""\n        elif not repl_lines:\n            sed_cmd = f""sed -i '{start_before},{end_before}d' {target_file}""\n        else:\n            escaped_lines = [_escape_single_quotes_for_sed(line) for line in repl_lines]\n            sed_payload = ""\n"".join(escaped_lines)\n            sed_cmd = f""sed -i '{start_before},{end_before}c\\\n{sed_payload}' {target_file}""\n        total_lines = len(after_state.splitlines())\n        center = (start_after + end_after) // 2\n        vp = _compute_viewport(total_lines, center, viewport_radius)\n        per_file_viewport[target_file] = vp\n        vstart, vend = vp\n        conversation_state.maybe_capture_file_contents(target_file, before_snapshot)\n        chained_cmd = f""{sed_cmd} && cat -n {target_file} | sed -n '{vstart},{vend}p'""\n        conversation_state.append_message({\n            ""from"": ""Assistant"",\n            ""value"": _fenced_block(""bash"", _clean_text(chained_cmd)),\n        })\n        viewport_output = _line_numbered_output(after_state, vstart, vend)\n        conversation_state.append_message({\n            ""from"": ""User"",\n            ""value"": f""<stdout>\n{viewport_output}\n</stdout>"",\n        })\n        pending_edits_before[target_file] = None\n        pending_edit_regions[target_file] = None\n\n    def _flush_all_pending_edits() -> None:\n        for fname in list(pending_edits_before.keys()):\n            _flush_pending_edit_for_file(fname)\n\n    for i in range(len(df)):\n        row = df.iloc[i]\n        file_path: str = row[""File""]\n        event_type = row[""Type""]\n\n        match event_type:\n            case ""tab"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                text = row[""Text""]\n                if pd.notna(text):\n                    content = str(text).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                    file_states[file_path] = content\n                    cmd = f""cat -n {file_path}""\n                    conversation_state.append_message({\n                        ""from"": ""Assistant"",\n                        ""value"": _fenced_block(""bash"", _clean_text(cmd)),\n                    })\n                    output = _line_numbered_output(content)\n                    conversation_state.append_message({\n                        ""from"": ""User"",\n                        ""value"": f""<stdout>\n{output}\n</stdout>"",\n                    })\n                    conversation_state.files_opened_in_conversation.add(file_path)\n                else:\n                    # File switch without content snapshot: show current viewport only\n                    content = file_states.get(file_path, """")\n                    total_lines = len(content.splitlines())\n                    vp = per_file_viewport.get(file_path)\n                    if not vp or vp[1] == 0:\n                        vp = _compute_viewport(total_lines, 1, viewport_radius)\n                        per_file_viewport[file_path] = vp\n                    if vp:\n                        vstart, vend = vp\n                        conversation_state.maybe_capture_file_contents(file_path, content)\n                        cmd = f""cat -n {file_path} | sed -n '{vstart},{vend}p'""\n                        conversation_state.append_message({\n                            ""from"": ""Assistant"",\n                            ""value"": _fenced_block(""bash"", _clean_text(cmd)),\n                        })\n                        viewport_output = _line_numbered_output(content, vstart, vend)\n                        conversation_state.append_message({\n                            ""from"": ""User"",\n                            ""value"": f""<stdout>\n{viewport_output}\n</stdout>"",\n                        })\n\n            case ""content"":\n                _flush_terminal_output_buffer()\n                offset = int(row[""RangeOffset""])\n                length = int(row[""RangeLength""])\n                new_text = row[""Text""]\n                before = file_states.get(file_path, """")\n                # Approximate current edit region in line space\n                new_text_str = str(new_text) if pd.notna(new_text) else """"\n                start_line_current = before[:offset].count(""\n"") + 1\n                deleted_conversation = before[offset:offset + length]\n                lines_added = new_text_str.count(""\n"")\n                lines_deleted = deleted_conversation.count(""\n"")\n                region_start = start_line_current\n                region_end = start_line_current + max(lines_added, lines_deleted, 0)\n                # Flush pending edits if this edit is far from the pending region\n                current_region = pending_edit_regions.get(file_path)\n                if current_region is not None:\n                    rstart, rend = current_region\n                    if region_start < (rstart - coalesce_radius) or region_start > (rend + coalesce_radius):\n                        _flush_pending_edit_for_file(file_path)\n                        current_region = None\n                after = _apply_change(before, offset, length, new_text)\n                if pending_edits_before.get(file_path) is None:\n                    pending_edits_before[file_path] = before\n                # Update/initialize region union\n                if current_region is None:\n                    pending_edit_regions[file_path] = (region_start, max(region_start, region_end))\n                else:\n                    rstart, rend = current_region\n                    pending_edit_regions[file_path] = (min(rstart, region_start), max(rend, region_end))\n                file_states[file_path] = after\n\n            case ""selection_command"" | ""selection_mouse"" | ""selection_keyboard"":\n                # During an edit burst (pending edits), suppress flush and viewport emissions\n                if pending_edits_before.get(file_path) is None:\n                    _flush_terminal_output_buffer()\n                else:\n                    # Skip emitting viewport while edits are pending to avoid per-keystroke sed/cat spam\n                    continue\n                offset = int(row[""RangeOffset""])\n                content = file_states.get(file_path, """")\n                total_lines = len(content.splitlines())\n                target_line = content[:offset].count(""\n"") + 1\n                vp = per_file_viewport.get(file_path)\n                should_emit = False\n                if not vp or vp[1] == 0:\n                    vp = _compute_viewport(total_lines, target_line, viewport_radius)\n                    per_file_viewport[file_path] = vp\n                    should_emit = True\n                else:\n                    vstart, vend = vp\n                    if target_line < vstart or target_line > vend:\n                        vp = _compute_viewport(total_lines, target_line, viewport_radius)\n                        per_file_viewport[file_path] = vp\n                        should_emit = True\n                if should_emit and vp:\n                    vstart, vend = vp\n                    conversation_state.maybe_capture_file_contents(file_path, content)\n                    cmd = f""cat -n {file_path} | sed -n '{vstart},{vend}p'""\n                    conversation_state.append_message({\n                        ""from"": ""Assistant"",\n                        ""value"": _fenced_block(""bash"", _clean_text(cmd)),\n                    })\n                    viewport_output = _line_numbered_output(content, vstart, vend)\n                    conversation_state.append_message({\n                        ""from"": ""User"",\n                        ""value"": f""<stdout>\n{viewport_output}\n</stdout>"",\n                    })\n\n            case ""terminal_command"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                command = row[""Text""]\n                command_str = str(command).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                conversation_state.append_message({\n                    ""from"": ""Assistant"",\n                    ""value"": _fenced_block(""bash"", _clean_text(command_str)),\n                })\n\n            case ""terminal_output"":\n                output = row[""Text""]\n                raw_output = str(output).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                terminal_output_buffer.append(raw_output)\n\n            case ""terminal_focus"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                # No-op for bash transcript; focus changes don't emit commands/output\n                pass\n\n            case ""git_branch_checkout"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                branch_info = row[""Text""]\n                branch_str = str(branch_info).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                cleaned = _clean_text(branch_str)\n                m = re.search(r""to '([^']+)'"", cleaned)\n                if not m:\n                    raise ValueError(f""Could not extract branch name from git checkout message: {cleaned}"")\n                branch_name = m.group(1).strip()\n                # Safe-quote branch if it contains special characters\n                if re.search(r""[^A-Za-z0-9._/\\-]"", branch_name):\n                    branch_name = ""'"" + branch_name.replace(""'"", ""'\""'\""'"") + ""'""\n                cmd = f""git checkout {branch_name}""\n                conversation_state.append_message({\n                    ""from"": ""Assistant"",\n                    ""value"": _fenced_block(""bash"", _clean_text(cmd)),\n                })\n\n            case _:\n                raise ValueError(f""Unknown event type: {event_type}"")\n\n    _flush_all_pending_edits()\n    _flush_terminal_output_buffer()\n    conversation_state.finalize_conversation()\n    return conversations, conversation_token_counts\n\n\n\ndef load_hf_csv(hf_path: str, split: str) -> Dataset:\n    loaded = load_dataset(hf_path, split=split)\n\n    assert isinstance(loaded, Dataset), ""Expected a Dataset from load_dataset""\n    return loaded\n\n\ndef _discover_local_sessions(root: Path) -> List[Path]:\n    # Recursively find all CSV files\n    paths: List[Path] = []\n    for p in root.rglob(""*.csv""):\n        if p.is_file():\n            paths.append(p)\n    paths.sort()\n    return paths",python,tab
+2,433,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"5:28:25 PM [info] Activating crowd-code\n5:28:25 PM [info] Recording started\n5:28:25 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,507,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"5:28:25 PM [info] Git repository found\n5:28:25 PM [info] Git provider initialized successfully\n5:28:25 PM [info] Initial git state: [object Object]\n",Log,content
+4,918,"TERMINAL",0,0,"bash",,terminal_focus
+5,920,"crowd-pilot/crowd_pilot/serialization_utils.py",0,0,"",python,tab
+6,920,"TERMINAL",0,0,"bash",,terminal_focus
+7,921,"TERMINAL",0,0,"bash",,terminal_focus
+8,921,"TERMINAL",0,0,"bash",,terminal_focus
+9,921,"TERMINAL",0,0,"bash",,terminal_focus
+10,921,"TERMINAL",0,0,"bash",,terminal_focus
+11,921,"TERMINAL",0,0,"bash",,terminal_focus
+12,921,"TERMINAL",0,0,"bash",,terminal_focus
+13,1115,"TERMINAL",0,0,"bash",,terminal_focus
+14,1116,"TERMINAL",0,0,"bash",,terminal_focus
+15,1305,"TERMINAL",0,0,"source /home/franz.srambical/crowd-pilot/.nemo_venv/bin/activate",,terminal_command
+16,1326,"TERMINAL",0,0,"source /home/franz.srambical/crowd-pilot/.nemo_venv/bin/activate",,terminal_command
+17,1327,"TERMINAL",0,0,"source /home/franz.srambical/crowd-pilot/.nemo_venv/bin/activate",,terminal_command
+18,1327,"TERMINAL",0,0,"source /home/franz.srambical/crowd-pilot/.nemo_venv/bin/activate",,terminal_command
+19,1327,"TERMINAL",0,0,"source /home/franz.srambical/crowd-pilot/.nemo_venv/bin/activate",,terminal_command
+20,1327,"TERMINAL",0,0,"source /home/franz.srambical/crowd-pilot/.nemo_venv/bin/activate",,terminal_command
+21,1327,"TERMINAL",0,0,"source /home/franz.srambical/crowd-pilot/.nemo_venv/bin/activate",,terminal_command
+22,1327,"TERMINAL",0,0,"source /home/franz.srambical/crowd-pilot/.nemo_venv/bin/activate",,terminal_command
+23,1345,"TERMINAL",0,0,"]633;C",,terminal_output
+24,1352,"TERMINAL",0,0,"]0;franz.srambical@hai-login1:~/crowd-pilot",,terminal_output
+25,1417,"TERMINAL",0,0,"source /home/franz.srambical/crowd-pilot/.nemo_venv/bin/activate",,terminal_command
+26,1428,"TERMINAL",0,0,"source /home/franz.srambical/crowd-pilot/.nemo_venv/bin/activate",,terminal_command
+27,1437,"TERMINAL",0,0,"]633;C]0;franz.srambical@hai-login1:~/crowd-pilot",,terminal_output
+28,12511,"TERMINAL",0,0,"squeue",,terminal_command
+29,12520,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  34970 mihir.maha interacti     1     2  R 2025-11-27T17:23:15 2025-11-27T17:23:15     5:22    2:00:00 hai003\r\n  34955 alfred.ngu interacti     1     8  R 2025-11-27T16:37:42 2025-11-27T16:37:42    50:55    2:00:00 hai003\r\n  34920   xiao.liu interacti     1   128  R 2025-11-27T07:17:03 2025-11-27T07:17:03 10:11:34   23:59:00 hai005\r\n  34917   xiao.liu interacti     1   128  R 2025-11-27T01:43:43 2025-11-27T01:43:43 15:44:54   23:59:00 hai006\r\n  34960 mihir.maha  standard     1    10  R 2025-11-27T17:12:37 2025-11-27T17:15:18    13:19      30:00 hai008\r\n  34964 nishant.ku  standard     1    32  R 2025-11-27T16:51:08 2025-11-27T16:51:08    37:29    6:00:00 hai007\r\n  34946 nick.lecht  standard     1    64  R 2025-11-27T15:29:49 2025-11-27T15:29:49  1:58:48   10:00:00 hai001\r\n  34939   xiao.liu  standard     1   128  R 2025-11-27T12:23:53 2025-11-27T13:50:03  3:38:34   23:59:00 hai004\r\n]0;franz.srambical@hai-login1:~/crowd-pilot",,terminal_output
+30,20218,"TERMINAL",0,0,"deactivate",,terminal_command
+31,29020,"TERMINAL",0,0,"source /home/franz.srambical/crowd-pilot/.venv/bin/activate",,terminal_command
+32,427924,"crowd-pilot/crowd_pilot/serialization_utils.py",21588,0,"",python,selection_command
+33,429372,"crowd-pilot/crowd_pilot/serialization_utils.py",0,0,"",python,selection_command
+34,430943,"crowd-pilot/crowd_pilot/serialize_dataset_nemo_json.py",0,0,"#!/usr/bin/env python3\n""""""\nCSV sessions -> JSONL (newline-delimited JSON) for NeMo 2.0 SFT function calling.\n\nEach session is converted to a multi-turn conversation where:\n- User role: terminal output (<stdout>...</stdout> content)\n- Assistant role: terminal commands (bash commands)\n""""""\n\nfrom __future__ import annotations\n\nimport argparse\nimport json\nimport os\nimport random\nfrom pathlib import Path\nfrom typing import List, Tuple, cast, Optional\nfrom dataclasses import dataclass\n\nimport pandas as pd\nfrom transformers import AutoTokenizer\n\nfrom serialization_utils import (\n    session_to_nemo_conversations,\n    _discover_local_sessions,\n)\n\n@dataclass\nclass SerializeConfig:\n    output_dir: str\n    max_tokens_per_conversation: int\n    max_tokens_per_message: int\n    min_conversation_messages: int\n    csv_root: Optional[str]\n    val_ratio: float\n    tokenizer_model: str\n\n\ndef to_nemo_jsonl(cfg: SerializeConfig) -> None:\n    """"""Convert CSV sessions to NeMo JSONL format.""""""\n    assert cfg.max_tokens_per_conversation > 0, ""max_tokens_per_conversation must be positive""\n    assert cfg.max_tokens_per_message > 0, ""max_tokens_per_message must be positive""\n    assert cfg.min_conversation_messages > 0, ""min_conversation_messages must be positive""\n    os.makedirs(cfg.output_dir, exist_ok=True)\n    \n    print(f""Loading tokenizer from {cfg.tokenizer_model}..."")\n    tokenizer = AutoTokenizer.from_pretrained(cfg.tokenizer_model)\n\n    required_cols = [""Sequence"", ""Time"", ""File"", ""RangeOffset"", ""RangeLength"", ""Text"", ""Language"", ""Type""]\n\n    session_dataframes: List[Tuple[pd.DataFrame, str]] = []\n    root = Path(cast(str, cfg.csv_root)).expanduser().resolve()\n    csv_files = _discover_local_sessions(root)\n    assert csv_files, f""No CSV files found under {root}""\n    \n    for csv_file in csv_files:\n        df = pd.read_csv(csv_file)\n        missing_local = [c for c in required_cols if c not in df.columns]\n        assert not missing_local, f""Missing required CSV columns in {csv_file}: {missing_local}""\n        session_dataframes.append((df, str(csv_file)))\n\n    random.seed(42)\n    random.shuffle(session_dataframes)\n    \n    total_sessions = len(session_dataframes)\n    val_count = int(total_sessions * cfg.val_ratio)\n    train_count = total_sessions - val_count\n\n    train_conversations = []\n    val_conversations = []\n    \n    message_counts: List[int] = []\n    token_counts: List[int] = []\n    conversations_written = 0\n\n    for i, (session_df, _) in enumerate(session_dataframes):\n        conversations, per_conversation_tokens = session_to_nemo_conversations(\n            session_df,\n            cfg.max_tokens_per_conversation,\n            max_tokens_per_message=cfg.max_tokens_per_message,\n            min_conversation_messages=cfg.min_conversation_messages,\n            tokenizer=tokenizer,\n        )\n\n        # Per-conversation statistics (for reporting)\n        per_conversation_messages = [len(conversation) for conversation in conversations]\n        \n        message_counts.extend(per_conversation_messages)\n        token_counts.extend(per_conversation_tokens)\n\n        for conversation in conversations:\n            record = {\n                ""mask"": ""User"",\n                ""system"": ""You are a helpful assistant that can interact multiple times with a computer shell to solve programming tasks.\nYour response must contain exactly ONE bash code block with ONE command (or commands connected with && or ||).\n\nFormat your response as shown in <format_example>.\n\n<format_example>\n```bash\nyour_command_here\n```\n</format_example>\n\nFailure to follow these rules will cause your response to be rejected."",\n                ""conversations"": conversation,\n            }\n\n            if i < train_count:\n                train_conversations.append(record)\n            else:\n                val_conversations.append(record)\n\n            conversations_written += 1\n\n    train_path = Path(cfg.output_dir) / ""training.jsonl""\n    with open(train_path, 'w', encoding='utf-8') as f:\n        for record in train_conversations:\n            f.write(json.dumps(record, ensure_ascii=False) + '\n')\n    \n    val_path = Path(cfg.output_dir) / ""validation.jsonl""\n    with open(val_path, 'w', encoding='utf-8') as f:\n        for record in val_conversations:\n            f.write(json.dumps(record, ensure_ascii=False) + '\n')\n\n    print(f""\n[summary]"")\n    print(f""  Total sessions processed: {total_sessions}"")\n    print(f""  Train conversations: {len(train_conversations)}"")\n    print(f""  Val conversations: {len(val_conversations)}"")\n    print(f""  Output: {cfg.output_dir}/{{training,validation}}.jsonl"")\n\n    total_messages = sum(message_counts)\n    total_tokens = sum(token_counts)\n    count = len(message_counts)\n\n    metadata = {\n        ""config"": {\n            ""csv_root"": cfg.csv_root,\n            ""output_dir"": cfg.output_dir,\n            ""min_conversation_messages"": cfg.min_conversation_messages,\n            ""val_ratio"": cfg.val_ratio,\n            ""max_tokens_per_conversation"": cfg.max_tokens_per_conversation,\n            ""max_tokens_per_message"": cfg.max_tokens_per_message,\n            ""tokenizer_model"": cfg.tokenizer_model,\n        },\n        ""counts"": {\n            ""total_sessions"": total_sessions,\n            ""train_conversations"": len(train_conversations),\n            ""val_conversations"": len(val_conversations),\n            ""conversations_written"": conversations_written,\n        },\n        ""stats"": {\n            ""messages"": {\n                ""total"": total_messages,\n                ""avg"": total_messages / count if count > 0 else 0,\n                ""min"": min(message_counts) if message_counts else 0,\n                ""max"": max(message_counts) if message_counts else 0,\n            },\n            ""tokens"": {\n                ""total"": total_tokens,\n                ""avg"": total_tokens / count if count > 0 else 0,\n                ""min"": min(token_counts) if token_counts else 0,\n                ""max"": max(token_counts) if token_counts else 0,\n            },\n        },\n        ""files"": {\n            ""train_path"": str(train_path),\n            ""val_path"": str(val_path),\n        },\n    }\n    metadata_path = Path(cfg.output_dir) / ""metadata.json""\n    with open(metadata_path, ""w"", encoding=""utf-8"") as mf:\n        json.dump(metadata, mf, ensure_ascii=False, indent=2)\n    print(f""  Metadata: {metadata_path}"")\n\n\ndef parse_args() -> SerializeConfig:\n    p = argparse.ArgumentParser(\n        description=""Serialize CSV sessions to JSONL for NeMo 2.0 SFT function calling""\n    )\n    p.add_argument(""--csv_root"", type=str, required=True, \n                   help=""Root directory containing per-session CSV files"")\n    p.add_argument(""--output_dir"", type=str, required=True, \n                   help=""Output directory for JSONL files"")\n    p.add_argument(""--min_conversation_messages"", type=int, default=5, \n                   help=""Minimum number of messages to keep a conversation chunk"")\n    p.add_argument(""--val_ratio"", type=float, default=0.10, \n                   help=""Fraction of sessions to route to validation [0,1)"")\n    p.add_argument(\n        ""--max_tokens_per_conversation"",\n        type=int,\n        default=8192,\n        help=""Maximum tokens per conversation chunk"",\n    )\n    p.add_argument(""--max_tokens_per_message"", type=int, default=2048, help=""Maximum tokens per message"")\n    p.add_argument(""--tokenizer_model"", type=str, required=True, help=""Path or name of the HuggingFace tokenizer model"")\n    \n    args = p.parse_args()\n    return SerializeConfig(\n        output_dir=args.output_dir,\n        max_tokens_per_conversation=args.max_tokens_per_conversation,\n        max_tokens_per_message=args.max_tokens_per_message,\n        min_conversation_messages=args.min_conversation_messages,\n        csv_root=(args.csv_root if args.csv_root else None),\n        val_ratio=args.val_ratio,\n        tokenizer_model=args.tokenizer_model,\n    )\n\n\ndef main() -> None:\n    cfg = parse_args()\n    to_nemo_jsonl(cfg)\n\n\nif __name__ == ""__main__"":\n    main()\n\n",python,tab
+35,433493,"crowd-pilot/crowd_pilot/serialize_dataset_nemo_json.py",3201,0,"",python,selection_command
+36,433923,"crowd-pilot/crowd_pilot/serialize_dataset_nemo_json.py",3207,0,"",python,selection_command
+37,434054,"crowd-pilot/crowd_pilot/serialize_dataset_nemo_json.py",3210,0,"",python,selection_command
+38,434290,"crowd-pilot/crowd_pilot/serialize_dataset_nemo_json.py",3211,0,"",python,selection_command
+39,434892,"crowd-pilot/crowd_pilot/serialize_dataset_nemo_json.py",3211,1,"Y",python,selection_command
+40,435158,"crowd-pilot/crowd_pilot/serialize_dataset_nemo_json.py",3211,3,"You",python,selection_command
+41,435348,"crowd-pilot/crowd_pilot/serialize_dataset_nemo_json.py",3211,7,"You are",python,selection_command
+42,436190,"crowd-pilot/crowd_pilot/serialize_dataset_nemo_json.py",3211,423,"You are a helpful assistant that can interact multiple times with a computer shell to solve programming tasks.\nYour response must contain exactly ONE bash code block with ONE command (or commands connected with && or ||).\n\nFormat your response as shown in <format_example>.\n\n<format_example>\n```bash\nyour_command_here\n```\n</format_example>\n\nFailure to follow these rules will cause your response to be rejected.""",python,selection_command
+43,437133,"crowd-pilot/crowd_pilot/serialize_dataset_nemo_json.py",3211,422,"You are a helpful assistant that can interact multiple times with a computer shell to solve programming tasks.\nYour response must contain exactly ONE bash code block with ONE command (or commands connected with && or ||).\n\nFormat your response as shown in <format_example>.\n\n<format_example>\n```bash\nyour_command_here\n```\n</format_example>\n\nFailure to follow these rules will cause your response to be rejected.",python,selection_command
+44,437481,"crowd-pilot/crowd_pilot/serialize_dataset_nemo_json.py",3632,0,"",python,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-b8ef8518-bf52-489e-b2a9-5e402fd02c471760857710761-2025_10_19-09.08.39.428/source.csv

@@ -0,0 +1,30 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,1,"crowd_code_player/replay_file.py",0,0,"import pandas as pd\nimport curses\nimport time\nimport argparse\n\ndef offset_to_yx(content, offset):\n    """"""Converts a 1D string offset to 2D (y, x) coordinates.""""""\n    # Ensure offset is within the bounds of the content length\n    offset = min(len(content), int(offset))\n    \n    # Find the line number by counting newlines before the offset\n    y = content.count('\n', 0, offset)\n    \n    # Find the column number by finding the last newline before the offset\n    last_newline_pos = content.rfind('\n', 0, offset)\n    if last_newline_pos == -1:\n        x = offset\n    else:\n        x = offset - last_newline_pos - 1\n        \n    return y, x\n\ndef apply_change(content, offset, length, new_text):\n    """"""Applies a text change to the content string.""""""\n    content = str(content)\n    new_text = str(new_text) if pd.notna(new_text) else """"\n    offset, length = int(offset), int(length)\n    \n    # Convert literal \n and \r characters to actual newlines and carriage returns\n    new_text = new_text.replace('\\n', '\n').replace('\\r', '\r')\n    \n    if offset > len(content):\n        content += ' ' * (offset - len(content)) # Pad if offset is out of bounds\n        \n    return content[:offset] + new_text + content[offset + length:]\n\ndef replay_trace(stdscr, filepath, speed_factor, long_pause_threshold=120000):\n    """"""Main function to replay the coding trace in the terminal.""""""\n    # --- Curses Setup ---\n    curses.curs_set(0) # We'll draw our own cursor\n    stdscr.nodelay(1)\n    curses.start_color()\n    curses.use_default_colors()\n    curses.init_pair(1, curses.COLOR_WHITE, -1) # For status bar\n    curses.init_pair(2, curses.COLOR_BLACK, curses.COLOR_WHITE) # For our cursor\n\n    # --- Data Loading ---\n    try:\n        df = pd.read_csv(filepath).sort_values('Time').reset_index(drop=True)\n    except FileNotFoundError:\n        print(f""Error: The file '{filepath}' was not found."")\n        return\n\n    # --- State Management ---\n    file_states = {}\n    scroll_states = {} # Tracks the top-line for each file's viewport\n    active_file = None\n    paused = False\n    \n    # --- Main Replay Loop ---\n    for i in range(len(df)):\n        # --- Handle User Input for Playback Control ---\n        key = stdscr.getch()\n        if key == ord('q'): break\n        if key == ord(' '): paused = not paused\n        if key == curses.KEY_UP: speed_factor = min(100, speed_factor * 1.5)\n        if key == curses.KEY_DOWN: speed_factor = max(0.1, speed_factor / 1.5)\n        \n        # Handle Paused State\n        if paused:\n            height, width = stdscr.getmaxyx()\n            stdscr.addstr(height - 1, 0, ""PAUSED"".ljust(width - 1), curses.A_REVERSE)\n            stdscr.refresh()\n            while paused:\n                time.sleep(0.1)\n                key = stdscr.getch()\n                if key == ord(' '): paused = False\n                elif key == ord('q'): return\n\n        # --- Process Event ---\n        event = df.iloc[i]\n        active_file = event['File']\n        \n        # Initialize state for new files\n        if active_file not in file_states:\n            file_states[active_file] = """"\n            scroll_states[active_file] = 0\n\n        \n        # Apply content change based on event type\n        if active_file == ""TERMINAL"":\n            # For terminal, just append text and add a newline\n            terminal_text = str(event['Text']) if pd.notna(event['Text']) else """"\n            # Convert literal \n and \r characters to actual newlines and carriage returns\n            terminal_text = terminal_text.replace('\\n', '\n').replace('\\r', '\r')\n            file_states[active_file] += terminal_text + '\n'\n        else:\n            file_states[active_file] = apply_change(\n                file_states[active_file], event['RangeOffset'], \n                event['RangeLength'], event['Text']\n            )\n            \n        # --- Calculate Cursor and Scrolling ---\n        content = file_states[active_file]\n        cursor_y, cursor_x = offset_to_yx(content, event['RangeOffset'])\n        scroll_y = scroll_states[active_file]\n        height, width = stdscr.getmaxyx()\n        visible_height = height - 2 # Account for status bars\n\n        # Adjust scroll to keep cursor in view\n        if active_file == ""TERMINAL"":\n            # For terminal, always scroll to bottom to show latest content\n            lines = content.split('\n')\n            total_lines = len(lines)\n            if total_lines > visible_height:\n                scroll_y = max(0, total_lines - visible_height)\n        else:\n            # For regular files, keep cursor in view\n            if cursor_y < scroll_y:\n                scroll_y = cursor_y\n            elif cursor_y >= scroll_y + visible_height:\n                scroll_y = cursor_y - visible_height + 1\n        \n        scroll_states[active_file] = scroll_y\n\n        # --- Render to Screen ---\n        stdscr.clear()\n        \n        # Display file content with scrolling\n        lines = content.split('\n')\n        for j in range(visible_height):\n            line_idx = scroll_y + j\n            if line_idx < len(lines):\n                stdscr.addstr(j, 0, lines[line_idx][:width - 1])\n        \n        # Draw our custom cursor\n        display_y = cursor_y - scroll_y\n        if 0 <= display_y < visible_height and 0 <= cursor_x < width:\n            # Ensure we don't try to draw on a non-existent character\n            line_len = len(lines[cursor_y]) if cursor_y < len(lines) else 0\n            char_to_draw_under = lines[cursor_y][cursor_x] if cursor_x < line_len else "" ""\n            stdscr.attron(curses.color_pair(2))\n            stdscr.addstr(display_y, cursor_x, char_to_draw_under)\n            stdscr.attroff(curses.color_pair(2))\n\n        # Status Bar\n        status_bar_text = f""File: {active_file} | Time: {event['Time']/1000:.1f}s | Event: {event['Type']} | Speed: {speed_factor:.1f}x""\n        stdscr.attron(curses.color_pair(1) | curses.A_REVERSE)\n        stdscr.addstr(height - 2, 0, status_bar_text.ljust(width - 1))\n        stdscr.attroff(curses.color_pair(1) | curses.A_REVERSE)\n\n        # Help Text\n        help_text = ""PAUSE/PLAY [space] | FASTER [↑] | SLOWER [↓] | QUIT [q]""\n        stdscr.addstr(height - 1, 0, help_text)\n\n        stdscr.refresh()\n\n        # --- Wait for Next Event ---\n        if i + 1 < len(df):\n            time_delta_ms = df.iloc[i+1]['Time'] - event['Time']\n            sleep_duration_s = max(0, time_delta_ms / 1000.0)\n            \n            # Check for long pauses\n            if time_delta_ms > long_pause_threshold:\n                # Display long pause message\n                height, width = stdscr.getmaxyx()\n                pause_message = ""Long pause detected. User might be googling, thinking or might have gone for a coffee...""\n                stdscr.addstr(height - 3, 0, pause_message.ljust(width - 1), curses.A_REVERSE)\n                stdscr.refresh()\n                time.sleep(1)  # Show message for 1 seconds\n                stdscr.clear()\n            else:\n                time.sleep(sleep_duration_s / speed_factor)\n\nif __name__ == ""__main__"":\n    parser = argparse.ArgumentParser(description=""Replay coding traces from a CSV file in the terminal."")\n    parser.add_argument(""filepath"", help=""The path to the source CSV file."")\n    parser.add_argument(""--speed"", type=float, default=20.0, help=""Initial playback speed multiplier."")\n    parser.add_argument(""--long_pause_threshold"", type=int, default=120000, help=""Threshold for long pause in milliseconds."")\n    args = parser.parse_args()\n\n    curses.wrapper(replay_trace, args.filepath, args.speed, args.long_pause_threshold)",python,tab
+2,70,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"9:08:39 AM [info] Activating crowd-code\n9:08:39 AM [info] Recording started\n9:08:39 AM [info] Initializing git provider using file system watchers...\n9:08:39 AM [info] Git repository found\n9:08:39 AM [info] Git provider initialized successfully\n",Log,tab
+3,207,"extension-output-pdoom-org.crowd-code-#1-crowd-code",245,0,"9:08:39 AM [info] Initial git state: [object Object]\n",Log,content
+4,13185962,"crowd_code_player/replay_file.py",0,0,"",python,tab
+5,13191919,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+6,13194346,"crowd_code_player/replay_file.py",0,0,"",python,tab
+7,34136506,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",0,0,"\n\documentclass{article} % For LaTeX2e\n\usepackage{iclr2026_conference,times}\n\n% Optional math commands from https://github.com/goodfeli/dlbook_notation.\n\input{math_commands.tex}\n\n\usepackage{hyperref}\n\usepackage{url}\n\n\n\title{Formatting Instructions for ICLR 2026 \\ Conference Submissions}\n\n% Authors must not appear in the submitted version. They should be hidden\n% as long as the \iclrfinalcopy macro remains commented out below.\n% Non-anonymous submissions will be rejected without review.\n\n\author{Antiquus S.~Hippocampus, Natalia Cerebro \& Amelie P. Amygdale \thanks{ Use footnote for providing further information\nabout author (webpage, alternative address)---\emph{not} for acknowledging\nfunding agencies.  Funding acknowledgements go at the end of the paper.} \\\nDepartment of Computer Science\\\nCranberry-Lemon University\\\nPittsburgh, PA 15213, USA \\\n\texttt{\{hippo,brain,jen\}@cs.cranberry-lemon.edu} \\\n\And\nJi Q. Ren \& Yevgeny LeNet \\\nDepartment of Computational Neuroscience \\\nUniversity of the Witwatersrand \\\nJoburg, South Africa \\\n\texttt{\{robot,net\}@wits.ac.za} \\\n\AND\nCoauthor \\\nAffiliation \\\nAddress \\\n\texttt{email}\n}\n\n% The \author macro works with any number of authors. There are two commands\n% used to separate the names and addresses of multiple authors: \And and \AND.\n%\n% Using \And between authors leaves it to \LaTeX{} to determine where to break\n% the lines. Using \AND forces a linebreak at that point. So, if \LaTeX{}\n% puts 3 of 4 authors names on the first line, and the last on the second\n% line, try using \AND instead of \And before the third author name.\n\n\newcommand{\fix}{\marginpar{FIX}}\n\newcommand{\new}{\marginpar{NEW}}\n\n%\iclrfinalcopy % Uncomment for camera-ready version, but NOT for submission.\n\begin{document}\n\n\n\maketitle\n\n\begin{abstract}\nThe abstract paragraph should be indented 1/2~inch (3~picas) on both left and\nright-hand margins. Use 10~point type, with a vertical spacing of 11~points.\nThe word \textsc{Abstract} must be centered, in small caps, and in point size 12. Two\nline spaces precede the abstract. The abstract must be limited to one\nparagraph.\n\end{abstract}\n\n\section{Submission of conference papers to ICLR 2026}\n\nICLR requires electronic submissions, processed by\n\url{https://openreview.net/}. See ICLR's website for more instructions.\n\nIf your paper is ultimately accepted, the statement {\tt\n  {\textbackslash}iclrfinalcopy} should be inserted to adjust the\nformat to the camera ready requirements.\n\nThe format for the submissions is a variant of the NeurIPS format.\nPlease read carefully the instructions below, and follow them\nfaithfully.\n\n\subsection{Style}\n\nPapers to be submitted to ICLR 2026 must be prepared according to the\ninstructions presented here.\n\n%% Please note that we have introduced automatic line number generation\n%% into the style file for \LaTeXe. This is to help reviewers\n%% refer to specific lines of the paper when they make their comments. Please do\n%% NOT refer to these line numbers in your paper as they will be removed from the\n%% style file for the final version of accepted papers.\n\nAuthors are required to use the ICLR \LaTeX{} style files obtainable at the\nICLR website. Please make sure you use the current files and\nnot previous versions. Tweaking the style files may be grounds for rejection.\n\n\subsection{Retrieval of style files}\n\nThe style files for ICLR and other conference information are available online at:\n\begin{center}\n   \url{http://www.iclr.cc/}\n\end{center}\nThe file \verb+iclr2026_conference.pdf+ contains these\ninstructions and illustrates the\nvarious formatting requirements your ICLR paper must satisfy.\nSubmissions must be made using \LaTeX{} and the style files\n\verb+iclr2026_conference.sty+ and \verb+iclr2026_conference.bst+ (to be used with \LaTeX{}2e). The file\n\verb+iclr2026_conference.tex+ may be used as a ``shell'' for writing your paper. All you\nhave to do is replace the author, title, abstract, and text of the paper with\nyour own.\n\nThe formatting instructions contained in these style files are summarized in\nsections \ref{gen_inst}, \ref{headings}, and \ref{others} below.\n\n\section{General formatting instructions}\n\label{gen_inst}\n\nThe text must be confined within a rectangle 5.5~inches (33~picas) wide and\n9~inches (54~picas) long. The left margin is 1.5~inch (9~picas).\nUse 10~point type with a vertical spacing of 11~points. Times New Roman is the\npreferred typeface throughout. Paragraphs are separated by 1/2~line space,\nwith no indentation.\n\nPaper title is 17~point, in small caps and left-aligned.\nAll pages should start at 1~inch (6~picas) from the top of the page.\n\nAuthors' names are\nset in boldface, and each name is placed above its corresponding\naddress. The lead author's name is to be listed first, and\nthe co-authors' names are set to follow. Authors sharing the\nsame address can be on the same line.\n\nPlease pay special attention to the instructions in section \ref{others}\nregarding figures, tables, acknowledgments, and references.\n\n\nThere will be a strict upper limit of \textbf{9 pages} for the main text of the initial submission, with unlimited additional pages for citations. This limit will be expanded to \textbf{10 pages} for rebuttal/camera ready.\n\n\section{Headings: first level}\n\label{headings}\n\nFirst level headings are in small caps,\nflush left and in point size 12. One line space before the first level\nheading and 1/2~line space after the first level heading.\n\n\subsection{Headings: second level}\n\nSecond level headings are in small caps,\nflush left and in point size 10. One line space before the second level\nheading and 1/2~line space after the second level heading.\n\n\subsubsection{Headings: third level}\n\nThird level headings are in small caps,\nflush left and in point size 10. One line space before the third level\nheading and 1/2~line space after the third level heading.\n\n\section{Citations, figures, tables, references}\n\label{others}\n\nThese instructions apply to everyone, regardless of the formatter being used.\n\n\subsection{Citations within the text}\n\nCitations within the text should be based on the \texttt{natbib} package\nand include the authors' last names and year (with the ``et~al.'' construct\nfor more than two authors). When the authors or the publication are\nincluded in the sentence, the citation should not be in parenthesis using \verb|\citet{}| (as\nin ``See \citet{Hinton06} for more information.''). Otherwise, the citation\nshould be in parenthesis using \verb|\citep{}| (as in ``Deep learning shows promise to make progress\ntowards AI~\citep{Bengio+chapter2007}.'').\n\nThe corresponding references are to be listed in alphabetical order of\nauthors, in the \textsc{References} section. As to the format of the\nreferences themselves, any style is acceptable as long as it is used\nconsistently.\n\n\subsection{Footnotes}\n\nIndicate footnotes with a number\footnote{Sample of the first footnote} in the\ntext. Place the footnotes at the bottom of the page on which they appear.\nPrecede the footnote with a horizontal rule of 2~inches\n(12~picas).\footnote{Sample of the second footnote}\n\n\subsection{Figures}\n\nAll artwork must be neat, clean, and legible. Lines should be dark\nenough for purposes of reproduction; art work should not be\nhand-drawn. The figure number and caption always appear after the\nfigure. Place one line space before the figure caption, and one line\nspace after the figure. The figure caption is lower case (except for\nfirst word and proper nouns); figures are numbered consecutively.\n\nMake sure the figure caption does not get separated from the figure.\nLeave sufficient space to avoid splitting the figure and figure caption.\n\nYou may use color figures.\nHowever, it is best for the\nfigure captions and the paper body to make sense if the paper is printed\neither in black/white or in color.\n\begin{figure}[h]\n\begin{center}\n%\framebox[4.0in]{$\;$}\n\fbox{\rule[-.5cm]{0cm}{4cm} \rule[-.5cm]{4cm}{0cm}}\n\end{center}\n\caption{Sample figure caption.}\n\end{figure}\n\n\subsection{Tables}\n\nAll tables must be centered, neat, clean and legible. Do not use hand-drawn\ntables. The table number and title always appear before the table. See\nTable~\ref{sample-table}.\n\nPlace one line space before the table title, one line space after the table\ntitle, and one line space after the table. The table title must be lower case\n(except for first word and proper nouns); tables are numbered consecutively.\n\n\begin{table}[t]\n\caption{Sample table title}\n\label{sample-table}\n\begin{center}\n\begin{tabular}{ll}\n\multicolumn{1}{c}{\bf PART}  &\multicolumn{1}{c}{\bf DESCRIPTION}\n\\ \hline \\\nDendrite         &Input terminal \\\nAxon             &Output terminal \\\nSoma             &Cell body (contains cell nucleus) \\\n\end{tabular}\n\end{center}\n\end{table}\n\n\section{Default Notation}\n\nIn an attempt to encourage standardized notation, we have included the\nnotation file from the textbook, \textit{Deep Learning}\n\cite{goodfellow2016deep} available at\n\url{https://github.com/goodfeli/dlbook_notation/}.  Use of this style\nis not required and can be disabled by commenting out\n\texttt{math\_commands.tex}.\n\n\n\centerline{\bf Numbers and Arrays}\n\bgroup\n\def\arraystretch{1.5}\n\begin{tabular}{p{1in}p{3.25in}}\n$\displaystyle a$ & A scalar (integer or real)\\\n$\displaystyle \va$ & A vector\\\n$\displaystyle \mA$ & A matrix\\\n$\displaystyle \tA$ & A tensor\\\n$\displaystyle \mI_n$ & Identity matrix with $n$ rows and $n$ columns\\\n$\displaystyle \mI$ & Identity matrix with dimensionality implied by context\\\n$\displaystyle \ve^{(i)}$ & Standard basis vector $[0,\dots,0,1,0,\dots,0]$ with a 1 at position $i$\\\n$\displaystyle \text{diag}(\va)$ & A square, diagonal matrix with diagonal entries given by $\va$\\\n$\displaystyle \ra$ & A scalar random variable\\\n$\displaystyle \rva$ & A vector-valued random variable\\\n$\displaystyle \rmA$ & A matrix-valued random variable\\\n\end{tabular}\n\egroup\n\vspace{0.25cm}\n\n\centerline{\bf Sets and Graphs}\n\bgroup\n\def\arraystretch{1.5}\n\n\begin{tabular}{p{1.25in}p{3.25in}}\n$\displaystyle \sA$ & A set\\\n$\displaystyle \R$ & The set of real numbers \\\n$\displaystyle \{0, 1\}$ & The set containing 0 and 1 \\\n$\displaystyle \{0, 1, \dots, n \}$ & The set of all integers between $0$ and $n$\\\n$\displaystyle [a, b]$ & The real interval including $a$ and $b$\\\n$\displaystyle (a, b]$ & The real interval excluding $a$ but including $b$\\\n$\displaystyle \sA \backslash \sB$ & Set subtraction, i.e., the set containing the elements of $\sA$ that are not in $\sB$\\\n$\displaystyle \gG$ & A graph\\\n$\displaystyle \parents_\gG(\ervx_i)$ & The parents of $\ervx_i$ in $\gG$\n\end{tabular}\n\vspace{0.25cm}\n\n\n\centerline{\bf Indexing}\n\bgroup\n\def\arraystretch{1.5}\n\n\begin{tabular}{p{1.25in}p{3.25in}}\n$\displaystyle \eva_i$ & Element $i$ of vector $\va$, with indexing starting at 1 \\\n$\displaystyle \eva_{-i}$ & All elements of vector $\va$ except for element $i$ \\\n$\displaystyle \emA_{i,j}$ & Element $i, j$ of matrix $\mA$ \\\n$\displaystyle \mA_{i, :}$ & Row $i$ of matrix $\mA$ \\\n$\displaystyle \mA_{:, i}$ & Column $i$ of matrix $\mA$ \\\n$\displaystyle \etA_{i, j, k}$ & Element $(i, j, k)$ of a 3-D tensor $\tA$\\\n$\displaystyle \tA_{:, :, i}$ & 2-D slice of a 3-D tensor\\\n$\displaystyle \erva_i$ & Element $i$ of the random vector $\rva$ \\\n\end{tabular}\n\egroup\n\vspace{0.25cm}\n\n\n\centerline{\bf Calculus}\n\bgroup\n\def\arraystretch{1.5}\n\begin{tabular}{p{1.25in}p{3.25in}}\n% NOTE: the [2ex] on the next line adds extra height to that row of the table.\n% Without that command, the fraction on the first line is too tall and collides\n% with the fraction on the second line.\n$\displaystyle\frac{d y} {d x}$ & Derivative of $y$ with respect to $x$\\ [2ex]\n$\displaystyle \frac{\partial y} {\partial x} $ & Partial derivative of $y$ with respect to $x$ \\\n$\displaystyle \nabla_\vx y $ & Gradient of $y$ with respect to $\vx$ \\\n$\displaystyle \nabla_\mX y $ & Matrix derivatives of $y$ with respect to $\mX$ \\\n$\displaystyle \nabla_\tX y $ & Tensor containing derivatives of $y$ with respect to $\tX$ \\\n$\displaystyle \frac{\partial f}{\partial \vx} $ & Jacobian matrix $\mJ \in \R^{m\times n}$ of $f: \R^n \rightarrow \R^m$\\\n$\displaystyle \nabla_\vx^2 f(\vx)\text{ or }\mH( f)(\vx)$ & The Hessian matrix of $f$ at input point $\vx$\\\n$\displaystyle \int f(\vx) d\vx $ & Definite integral over the entire domain of $\vx$ \\\n$\displaystyle \int_\sS f(\vx) d\vx$ & Definite integral with respect to $\vx$ over the set $\sS$ \\\n\end{tabular}\n\egroup\n\vspace{0.25cm}\n\n\centerline{\bf Probability and Information Theory}\n\bgroup\n\def\arraystretch{1.5}\n\begin{tabular}{p{1.25in}p{3.25in}}\n$\displaystyle P(\ra)$ & A probability distribution over a discrete variable\\\n$\displaystyle p(\ra)$ & A probability distribution over a continuous variable, or over\na variable whose type has not been specified\\\n$\displaystyle \ra \sim P$ & Random variable $\ra$ has distribution $P$\\% so thing on left of \sim should always be a random variable, with name beginning with \r\n$\displaystyle  \E_{\rx\sim P} [ f(x) ]\text{ or } \E f(x)$ & Expectation of $f(x)$ with respect to $P(\rx)$ \\\n$\displaystyle \Var(f(x)) $ &  Variance of $f(x)$ under $P(\rx)$ \\\n$\displaystyle \Cov(f(x),g(x)) $ & Covariance of $f(x)$ and $g(x)$ under $P(\rx)$\\\n$\displaystyle H(\rx) $ & Shannon entropy of the random variable $\rx$\\\n$\displaystyle \KL ( P \Vert Q ) $ & Kullback-Leibler divergence of P and Q \\\n$\displaystyle \mathcal{N} ( \vx ; \vmu , \mSigma)$ & Gaussian distribution %\nover $\vx$ with mean $\vmu$ and covariance $\mSigma$ \\\n\end{tabular}\n\egroup\n\vspace{0.25cm}\n\n\centerline{\bf Functions}\n\bgroup\n\def\arraystretch{1.5}\n\begin{tabular}{p{1.25in}p{3.25in}}\n$\displaystyle f: \sA \rightarrow \sB$ & The function $f$ with domain $\sA$ and range $\sB$\\\n$\displaystyle f \circ g $ & Composition of the functions $f$ and $g$ \\\n  $\displaystyle f(\vx ; \vtheta) $ & A function of $\vx$ parametrized by $\vtheta$.\n  (Sometimes we write $f(\vx)$ and omit the argument $\vtheta$ to lighten notation) \\\n$\displaystyle \log x$ & Natural logarithm of $x$ \\\n$\displaystyle \sigma(x)$ & Logistic sigmoid, $\displaystyle \frac{1} {1 + \exp(-x)}$ \\\n$\displaystyle \zeta(x)$ & Softplus, $\log(1 + \exp(x))$ \\\n$\displaystyle || \vx ||_p $ & $\normlp$ norm of $\vx$ \\\n$\displaystyle || \vx || $ & $\normltwo$ norm of $\vx$ \\\n$\displaystyle x^+$ & Positive part of $x$, i.e., $\max(0,x)$\\\n$\displaystyle \1_\mathrm{condition}$ & is 1 if the condition is true, 0 otherwise\\\n\end{tabular}\n\egroup\n\vspace{0.25cm}\n\n\n\n\section{Final instructions}\nDo not change any aspects of the formatting parameters in the style files.\nIn particular, do not modify the width or length of the rectangle the text\nshould fit into, and do not change font sizes (except perhaps in the\n\textsc{References} section; see below). Please note that pages should be\nnumbered.\n\n\section{Preparing PostScript or PDF files}\n\nPlease prepare PostScript or PDF files with paper size ``US Letter'', and\nnot, for example, ``A4''. The -t\nletter option on dvips will produce US Letter files.\n\nConsider directly generating PDF files using \verb+pdflatex+\n(especially if you are a MiKTeX user).\nPDF figures must be substituted for EPS figures, however.\n\nOtherwise, please generate your PostScript and PDF files with the following commands:\n\begin{verbatim}\ndvips mypaper.dvi -t letter -Ppdf -G0 -o mypaper.ps\nps2pdf mypaper.ps mypaper.pdf\n\end{verbatim}\n\n\subsection{Margins in LaTeX}\n\nMost of the margin problems come from figures positioned by hand using\n\verb+\special+ or other commands. We suggest using the command\n\verb+\includegraphics+\nfrom the graphicx package. Always specify the figure width as a multiple of\nthe line width as in the example below using .eps graphics\n\begin{verbatim}\n   \usepackage[dvips]{graphicx} ...\n   \includegraphics[width=0.8\linewidth]{myfile.eps}\n\end{verbatim}\nor % Apr 2009 addition\n\begin{verbatim}\n   \usepackage[pdftex]{graphicx} ...\n   \includegraphics[width=0.8\linewidth]{myfile.pdf}\n\end{verbatim}\nfor .pdf graphics.\nSee section~4.4 in the graphics bundle documentation (\url{http://www.ctan.org/tex-archive/macros/latex/required/graphics/grfguide.ps})\n\nA number of width problems arise when LaTeX cannot properly hyphenate a\nline. Please give LaTeX hyphenation hints using the \verb+\-+ command.\n\n\subsubsection*{Author Contributions}\nIf you'd like to, you may include  a section for author contributions as is done\nin many journals. This is optional and at the discretion of the authors.\n\n\subsubsection*{Acknowledgments}\nUse unnumbered third level headings for the acknowledgments. All\nacknowledgments, including those to funding agencies, go at the end of the paper.\n\n\n\bibliography{iclr2026_conference}\n\bibliographystyle{iclr2026_conference}\n\n\appendix\n\section{Appendix}\nYou may include other additional sections here.\n\n\n\end{document}\n",latex,tab
+8,34141951,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",221,0,"",latex,selection_mouse
+9,34148743,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",222,0,"",latex,selection_command
+10,34148955,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",294,0,"",latex,selection_command
+11,34149125,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",295,0,"",latex,selection_command
+12,34149319,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",369,0,"",latex,selection_command
+13,34149487,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",436,0,"",latex,selection_command
+14,34149748,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",497,0,"",latex,selection_command
+15,34150077,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",498,0,"",latex,selection_command
+16,34150584,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",625,0,"",latex,selection_command
+17,34150929,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",498,0,"",latex,selection_command
+18,34152432,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",498,126,"\author{Antiquus S.~Hippocampus, Natalia Cerebro \& Amelie P. Amygdale \thanks{ Use footnote for providing further information",latex,selection_command
+19,34153871,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",498,0,"",latex,selection_command
+20,34154430,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",625,0,"",latex,selection_command
+21,34154681,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",700,0,"",latex,selection_command
+22,34154720,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",776,0,"",latex,selection_command
+23,34154750,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",809,0,"",latex,selection_command
+24,34154915,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",838,0,"",latex,selection_command
+25,34155072,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",867,0,"",latex,selection_command
+26,34155589,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",922,0,"",latex,selection_command
+27,34155796,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",927,0,"",latex,selection_command
+28,34156131,"/Users/franzsrambical/Downloads/iclr2026 2/iclr2026_conference.tex",922,0,"",latex,selection_command
+29,34188545,"crowd_code_player/replay_file.py",0,0,"",python,tab

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-b9325173-6eb8-4347-ba97-77b482c6beaf1760621262730-2025_10_16-15.27.51.183/source.csv

@@ -0,0 +1,16 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"slurm/dev/franz/berlin/coinrun/sample/maskgit/sample_mila_submission_case_study_vanilla.sh",0,0,"#!/usr/bin/env bash\n\n#SBATCH --nodes=1\n#SBATCH --ntasks-per-node=1\n#SBATCH --time=24:00:00\n#SBATCH --cpus-per-task=8\n#SBATCH --gres=gpu:1\n#SBATCH --output=/fast/project/HFMI_SynergyUnit/jafar_ws/logs/franz/coinrun/dynamics_sample/%x_%j.log\n#SBATCH --error=/fast/project/HFMI_SynergyUnit/jafar_ws/logs/franz/coinrun/dynamics_sample/%x_%j.log\n#SBATCH --job-name=coinrun_sample_maskgit_mila_submission_case_study_vanilla\n\n# Activate virtual environment\nsource .venv/bin/activate\n\narray_records_dir=""/fast/project/HFMI_SynergyUnit/jafar_ws/data/coinrun/array_records_10M_npy_arr_rec/array_record/test""\nCHECKPOINT_PATH=""/fast/project/HFMI_SynergyUnit/jafar_ws/checkpoints/coinrun/dynamics/dynamics_case_study_dataset_10M_30031""\n\ncurrent_branch=$(git rev-parse --abbrev-ref HEAD)\nif [ ""$current_branch"" != ""main"" ]; then\n    echo ""This script must be run from the main branch. Current branch is $current_branch. Exiting.""\n    exit 1\nfi\n\necho ""Sampling from checkpoint: $CHECKPOINT_PATH""\n\nsrun python jasmine/sample.py \\n    --seed=1 \\n    --maskgit_steps=1 \\n    --tokenizer_ffn_dim=512 \\n    --tokenizer_num_blocks=8 \\n    --dyna_ffn_dim=512 \\n    --dyna_num_blocks=12 \\n    --output_dir=gifs/dynamics_case_study_dataset_10M_vanilla \\n    --checkpoint $CHECKPOINT_PATH \\n    --data_dir=$array_records_dir \\n    --seq_len=16 \\n    --batch_size=32 \\n    --patch_size=4 \\n    --start_frame=4 \\n    --image_height=64 \\n    --image_width=64 \\n    --dyna_type=maskgit\n",shellscript,tab
+2,269,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"3:27:51 PM [info] Activating crowd-code\n3:27:51 PM [info] Recording started\n3:27:51 PM [info] Initializing git provider using file system watchers...\n3:27:51 PM [info] Git repository found\n3:27:51 PM [info] Git provider initialized successfully\n3:27:51 PM [info] Initial git state: [object Object]\n",Log,tab
+3,3151,"TERMINAL",0,0,"",,terminal_command
+4,9914,"TERMINAL",0,0,"",,terminal_command
+5,10731,"slurm/dev/franz/berlin/coinrun/sample/maskgit/sample_mila_submission_case_study_vanilla.sh",0,0,"",shellscript,tab
+6,20993,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+7,22957,"TERMINAL",0,0,"",,terminal_focus
+8,22958,"slurm/dev/franz/berlin/coinrun/sample/maskgit/sample_mila_submission_case_study_vanilla.sh",0,0,"",shellscript,tab
+9,23471,"TERMINAL",0,0,"source /home/franz.srambical/jafar/.venv/bin/activate",,terminal_command
+10,23483,"TERMINAL",0,0,"]633;C]0;franz.srambical@hai-login2:~/jafar",,terminal_output
+11,31573,"TERMINAL",0,0,"sacct --format=""JobID%15,JobName%30,Partition%16,AllocCPUS%3,State%12,Elapsed%10,Timelimit%10"" --starttime $(date -d ""last month"" +%Y-%m-%d) | grep -vE ""*.batch|*.extern|*.inter|bash|python|CANCELLED|echo""",,terminal_command
+12,31642,"TERMINAL",0,0,"]633;C          JobID                        JobName        Partition All        State    Elapsed  Timelimit \r\n--------------- ------------------------------ ---------------- --- ------------ ---------- ---------- \r\n          29284                           wrap         standard 128    COMPLETED   10:09:39 1-00:00:00 \r\n          29399            generate_atari_data         standard  16    COMPLETED   01:03:04 1-00:00:00 \r\n          29462            generate_atari_data         standard 128    COMPLETED   15:12:02 1-00:00:00 \r\n          29473 generate_minatar_breakout_data         standard 128    COMPLETED   02:02:11 1-00:00:00 \r\n          29671 generate_atari_breakout_no_st+         standard 128    COMPLETED   08:39:39 1-00:00:00 \r\n          29672 generate_atari_pong_no_sticky+         standard 128    COMPLETED   08:41:40 1-00:00:00 \r\n          29673 generate_atari_alien_no_stick+         standard 128       FAILED   00:00:14 1-00:00:00 \r\n          29674 generate_atari_amidar_no_stic+         standard 128       FAILED   00:00:14 1-00:00:00 \r\n          29675 generate_atari_assault_no_sti+         standard 128       FAILED   00:00:20 1-00:00:00 \r\n          29676 generate_atari_asterix_no_sti+         standard 128       FAILED   00:00:10 1-00:00:00 \r\n          29677 generate_atari_bank_heist_no_+         standard 128       FAILED   00:00:10 1-00:00:00 \r\n          29678 generate_atari_battle_zone_no+         standard 128       FAILED   00:00:10 1-00:00:00 \r\n          29679 generate_atari_boxing_no_stic+         standard 128       FAILED   00:00:10 1-00:00:00 \r\n          29680 generate_atari_chopper_comman+         standard 128       FAILED   00:00:10 1-00:00:00 \r\n          29681 generate_atari_crazy_climber_+         standard 128       FAILED   00:00:10 1-00:00:00 \r\n          29682 generate_atari_demon_attack_n+         standard 128       FAILED   00:00:09 1-00:00:00 \r\n          29683 generate_atari_alien_no_stick+         standard 128    COMPLETED   09:46:52 1-00:00:00 \r\n          29685 generate_atari_amidar_no_stic+         standard 128    COMPLETED   09:21:44 1-00:00:00 \r\n          29686 generate_atari_assault_no_sti+         standard 128    COMPLETED   08:32:12 1-00:00:00 \r\n          29687 generate_atari_asterix_no_sti+         standard 128    COMPLETED   08:51:17 1-00:00:00 \r\n          29688 generate_atari_bank_heist_no_+         standard 128    COMPLETED   09:02:48 1-00:00:00 \r\n          29689 generate_atari_battle_zone_no+         standard 128    COMPLETED   09:20:54 1-00:00:00 \r\n          29690 generate_atari_boxing_no_stic+         standard 128    COMPLETED   09:18:27 1-00:00:00 \r\n          29691 generate_atari_chopper_comman+         standard 128    COMPLETED   09:17:26 1-00:00:00 \r\n          29692 generate_atari_crazy_climber_+         standard 128    COMPLETED   08:49:47 1-00:00:00 \r\n          29693 generate_atari_demon_attack_n+         standard 128    COMPLETED   08:16:52 1-00:00:00 \r\n          29734 tokenizer_coinrun_mila_submis+         standard  16       FAILED   00:00:00 1-00:00:00 \r\n          29736 tokenizer_coinrun_mila_submis+         standard  16    COMPLETED   03:58:28 1-00:00:00 \r\n          29737    lam_coinrun_mila_submission         standard  16    COMPLETED   02:31:23 1-00:00:00 \r\n          29738 lam_coinrun_mila_submission_n+         standard  16    COMPLETED   02:35:49 1-00:00:00 \r\n          29759 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   06:46:34 1-00:00:00 \r\n          29760 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   04:02:30 1-00:00:00 \r\n          29761 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   05:09:04 1-00:00:00 \r\n          29762 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   05:22:21 1-00:00:00 \r\n          29763 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   05:11:06 1-00:00:00 \r\n          29764 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:21 1-00:00:00 \r\n          29765 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:21 1-00:00:00 \r\n          29766 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   05:35:34 1-00:00:00 \r\n          29771 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   14:03:09 1-00:00:00 \r\n          29772 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   14:47:34 1-00:00:00 \r\n          29773 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   04:36:48 1-00:00:00 \r\n          29804 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:01 1-00:00:00 \r\n          29805 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:00 1-00:00:00 \r\n          29807 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   14:44:21 1-00:00:00 \r\n          29808 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   04:29:03 1-00:00:00 \r\n          29845 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   04:24:16 1-00:00:00 \r\n          29898 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:01 1-00:00:00 \r\n          29899 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   03:25:34 1-00:00:00 \r\n          29906 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:01:16 1-00:00:00 \r\n          29923 tokenizer_coinrun_mila_submis+         standard  16       FAILED   02:22:42 1-00:00:00 \r\n          29924    lam_coinrun_mila_submission         standard  16       FAILED   02:21:54 1-00:00:00 \r\n          29927 tokenizer_coinrun_mila_submis+         standard  16       FAILED   01:04:49 1-00:00:00 \r\n          29928    lam_coinrun_mila_submission         standard  16       FAILED   01:04:47 1-00:00:00 \r\n          29949 tokenizer_coinrun_mila_submis+         standard  16    COMPLETED   03:48:24 1-00:00:00 \r\n          29950    lam_coinrun_mila_submission         standard  16    COMPLETED   02:20:48 1-00:00:00 \r\n          29959 lam_coinrun_mila_submission_n+         standard  16    COMPLETED   02:24:53 1-00:00:00 \r\n          29973 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   05:01:33 1-00:00:00 \r\n          29974 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   04:59:19 1-00:00:00 \r\n          29976 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   06:33:08 1-00:00:00 \r\n          29977 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   03:45:18 1-00:00:00 \r\n          29978 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   05:03:54 1-00:00:00 \r\n          29985 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   02:28:50 1-00:00:00 \r\n          29986 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   03:13:31 1-00:00:00 \r\n          29987 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   02:36:47 1-00:00:00 \r\n          29988 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:01 1-00:00:00 \r\n          29989 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:01 1-00:00:00 \r\n          29990 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   02:34:01 1-00:00:00 \r\n          29991 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:01 1-00:00:00 \r\n          29995 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:45 1-00:00:00 \r\n          29996 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   19:46:35 1-00:00:00 \r\n          30047      tokenizer_atari_alien_dev         standard  16       FAILED   00:00:00 1-00:00:00 \r\n          30048      tokenizer_atari_alien_dev         standard  16       FAILED   00:00:32 1-00:00:00 \r\n          30075           lam_atari_amidar_dev         standard  16    COMPLETED   03:53:54 1-00:00:00 \r\n          30076          lam_atari_assault_dev         standard  16    COMPLETED   03:52:52 1-00:00:00 \r\n          30077          lam_atari_asterix_dev         standard  16    COMPLETED   03:48:53 1-00:00:00 \r\n          30078       lam_atari_bank_heist_dev         standard  16    COMPLETED   02:58:29 1-00:00:00 \r\n          30079      lam_atari_battle_zone_dev         standard  16    COMPLETED   03:54:36 1-00:00:00 \r\n          30080           lam_atari_boxing_dev         standard  16    COMPLETED   02:59:00 1-00:00:00 \r\n          30081         lam_atari_breakout_dev         standard  16    COMPLETED   04:05:11 1-00:00:00 \r\n          30082  lam_atari_chopper_command_dev         standard  16       FAILED   00:00:20 1-00:00:00 \r\n          30083    lam_atari_crazy_climber_dev         standard  16    COMPLETED   04:08:20 1-00:00:00 \r\n          30084     lam_atari_demon_attack_dev         standard  16    COMPLETED   04:09:50 1-00:00:00 \r\n          30085             lam_atari_pong_dev         standard  16    COMPLETED   04:09:13 1-00:00:00 \r\n          30092      tokenizer_atari_alien_dev         standard  16    COMPLETED   02:17:47 1-00:00:00 \r\n          30093     tokenizer_atari_amidar_dev         standard  16    COMPLETED   01:48:07 1-00:00:00 \r\n          30094    tokenizer_atari_assault_dev         standard  16    COMPLETED   03:05:43 1-00:00:00 \r\n          30095    tokenizer_atari_asterix_dev         standard  16    COMPLETED   02:03:04 1-00:00:00 \r\n          30096 tokenizer_atari_bank_heist_dev         standard  16    COMPLETED   01:27:56 1-00:00:00 \r\n          30097 tokenizer_atari_battle_zone_d+         standard  16    COMPLETED   01:13:24 1-00:00:00 \r\n          30098     tokenizer_atari_boxing_dev         standard  16    COMPLETED   01:37:54 1-00:00:00 \r\n          30099   tokenizer_atari_breakout_dev         standard  16    COMPLETED   01:49:10 1-00:00:00 \r\n          30100 tokenizer_atari_chopper_comma+         standard  16    COMPLETED   01:16:14 1-00:00:00 \r\n          30101 tokenizer_atari_crazy_climber+         standard  16    COMPLETED   03:19:30 1-00:00:00 \r\n          30102 tokenizer_atari_demon_attack_+         standard  16    COMPLETED   01:57:06 1-00:00:00 \r\n          30103       tokenizer_atari_pong_dev         standard  16    COMPLETED   01:30:45 1-00:00:00 \r\n          30117       dynamics_atari_alien_dev         standard  16    COMPLETED   06:10:49 1-00:00:00 \r\n          30118      dynamics_atari_amidar_dev         standard  16    COMPLETED   05:23:49 1-00:00:00 \r\n          30119     dynamics_atari_assault_dev         standard  16    COMPLETED   06:20:45 1-00:00:00 \r\n          30120     dynamics_atari_asterix_dev         standard  16    COMPLETED   05:27:53 1-00:00:00 \r\n          30121  dynamics_atari_bank_heist_dev         standard  16    COMPLETED   05:28:06 1-00:00:00 \r\n          30122 dynamics_atari_battle_zone_dev         standard  16    COMPLETED   06:11:44 1-00:00:00 \r\n          30123      dynamics_atari_boxing_dev         standard  16    COMPLETED   06:14:07 1-00:00:00 \r\n          30124    dynamics_atari_breakout_dev         standard  16    COMPLETED   06:09:47 1-00:00:00 \r\n          30125 dynamics_atari_chopper_comman+         standard  16    COMPLETED   06:12:46 1-00:00:00 \r\n          30126        dynamics_atari_pong_dev         standard  16    COMPLETED   06:19:33 1-00:00:00 \r\n          30127 dynamics_atari_demon_attack_d+         standard  16    COMPLETED   06:16:14 1-00:00:00 \r\n          30130 tokenizer_atari_alien_dev_lr_+         standard  16    COMPLETED   03:07:39 1-00:00:00 \r\n          30131 tokenizer_atari_amidar_dev_lr+         standard  16    COMPLETED   02:13:25 1-00:00:00 \r\n          30132 tokenizer_atari_assault_dev_l+         standard  16    COMPLETED   03:28:04 1-00:00:00 \r\n          30133 tokenizer_atari_asterix_dev_l+         standard  16    COMPLETED   04:04:16 1-00:00:00 \r\n          30134 tokenizer_atari_bank_heist_de+         standard  16    COMPLETED   01:33:50 1-00:00:00 \r\n          30135 tokenizer_atari_battle_zone_d+         standard  16    COMPLETED   01:44:19 1-00:00:00 \r\n          30136 tokenizer_atari_boxing_dev_lr+         standard  16    COMPLETED   01:58:10 1-00:00:00 \r\n          30137 tokenizer_atari_breakout_dev_+         standard  16    COMPLETED   01:52:02 1-00:00:00 \r\n          30138 tokenizer_atari_chopper_comma+         standard  16    COMPLETED   01:28:25 1-00:00:00 \r\n          30139 tokenizer_atari_crazy_climber+         standard  16    COMPLETED   04:37:01 1-00:00:00 \r\n          30140 tokenizer_atari_demon_attack_+         standard  16    COMPLETED   03:58:56 1-00:00:00 \r\n          30141 tokenizer_atari_pong_dev_lr_3+         standard  16    COMPLETED   01:46:00 1-00:00:00 \r\n          30143 tokenizer_atari_alien_dev_lr_+         standard  16       FAILED   00:00:25 1-00:00:00 \r\n          30144 tokenizer_atari_amidar_dev_lr+         standard  16       FAILED   00:00:32 1-00:00:00 \r\n          30145 tokenizer_atari_assault_dev_l+         standard  16       FAILED   00:00:33 1-00:00:00 \r\n          30146 tokenizer_atari_asterix_dev_l+         standard  16       FAILED   00:00:34 1-00:00:00 \r\n          30147 tokenizer_atari_bank_heist_de+         standard  16       FAILED   00:00:41 1-00:00:00 \r\n          30148 tokenizer_atari_battle_zone_d+         standard  16       FAILED   00:00:38 1-00:00:00 \r\n          30149 tokenizer_atari_boxing_dev_lr+         standard  16       FAILED   00:00:30 1-00:00:00 \r\n          30150 tokenizer_atari_breakout_dev_+         standard  16       FAILED   00:00:29 1-00:00:00 \r\n          30151 tokenizer_atari_chopper_comma+         standard  16       FAILED   00:00:31 1-00:00:00 \r\n          30152 tokenizer_atari_crazy_climber+         standard  16       FAILED   00:00:34 1-00:00:00 \r\n          30153 tokenizer_atari_demon_attack_+         standard  16       FAILED   00:00:37 1-00:00:00 \r\n          30154 tokenizer_atari_pong_dev_lr_3+         standard  16       FAILED   00:00:32 1-00:00:00 \r\n          30155 tokenizer_atari_alien_dev_lr_+         standard  16       FAILED   00:00:43 1-00:00:00 \r\n          30156 tokenizer_atari_amidar_dev_lr+         standard  16       FAILED   00:00:36 1-00:00:00 \r\n          30157 tokenizer_atari_assault_dev_l+         standard  16       FAILED   00:00:28 1-00:00:00 \r\n          30158 tokenizer_atari_asterix_dev_l+         standard  16       FAILED   00:00:37 1-00:00:00 \r\n          30159 tokenizer_atari_bank_heist_de+         standard  16       FAILED   00:00:36 1-00:00:00 \r\n          30160 tokenizer_atari_battle_zone_d+         standard  16       FAILED   00:00:32 1-00:00:00 \r\n          30161 tokenizer_atari_boxing_dev_lr+         standard  16       FAILED   00:00:30 1-00:00:00 \r\n          30162 tokenizer_atari_breakout_dev_+         standard  16       FAILED   00:00:41 1-00:00:00 \r\n          30163 tokenizer_atari_chopper_comma+         standard  16       FAILED   00:00:21 1-00:00:00 \r\n          30164 tokenizer_atari_crazy_climber+         standard  16       FAILED   00:00:28 1-00:00:00 \r\n          30165 tokenizer_atari_demon_attack_+         standard  16       FAILED   00:00:38 1-00:00:00 \r\n          30166 tokenizer_atari_pong_dev_lr_3+         standard  16       FAILED   00:00:27 1-00:00:00 \r\n          30167 tokenizer_atari_alien_dev_lr_+         standard  16       FAILED   00:00:29 1-00:00:00 \r\n          30168 tokenizer_atari_amidar_dev_lr+         standard  16       FAILED   00:00:42 1-00:00:00 \r\n          30169 tokenizer_atari_assault_dev_l+         standard  16       FAILED   00:00:37 1-00:00:00 \r\n          30170 tokenizer_atari_asterix_dev_l+         standard  16       FAILED   00:00:33 1-00:00:00 \r\n          30171 tokenizer_atari_bank_heist_de+         standard  16       FAILED   00:00:35 1-00:00:00 \r\n          30172 tokenizer_atari_battle_zone_d+         standard  16       FAILED   00:00:38 1-00:00:00 \r\n          30173 tokenizer_atari_boxing_dev_lr+         standard  16       FAILED   00:00:26 1-00:00:00 \r\n          30174 tokenizer_atari_breakout_dev_+         standard  16       FAILED   00:00:30 1-00:00:00 \r\n          30175 tokenizer_atari_chopper_comma+         standard  16       FAILED   00:00:37 1-00:00:00 \r\n          30176 tokenizer_atari_crazy_climber+         standard  16       FAILED   00:00:32 1-00:00:00 \r\n          30177 tokenizer_atari_demon_attack_+         standard  16       FAILED   00:00:43 1-00:00:00 \r\n          30178 tokenizer_atari_pong_dev_lr_3+         standard  16       FAILED   00:00:33 1-00:00:00 \r\n          30193    doom_dataset_generation_10m         standard  16       FAILED   00:00:06 1-00:00:00 \r\n          30194    doom_dataset_generation_10m         standard  16       FAILED   00:00:19 1-00:00:00 \r\n          30195    doom_dataset_generation_10m         standard  16       FAILED   00:00:09 1-00:00:00 \r\n          30196 tokenizer_atari_alien_dev_lr_+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30197 tokenizer_atari_amidar_dev_lr+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30198 tokenizer_atari_assault_dev_l+         standard  16       FAILED   00:00:01 1-00:00:00 \r\n          30199 tokenizer_atari_asterix_dev_l+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30200 tokenizer_atari_bank_heist_de+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30201 tokenizer_atari_battle_zone_d+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30202 tokenizer_atari_boxing_dev_lr+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30203 tokenizer_atari_breakout_dev_+         standard  16       FAILED   00:00:01 1-00:00:00 \r\n          30204 tokenizer_atari_chopper_comma+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30205 tokenizer_atari_crazy_climber+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30206 tokenizer_atari_demon_attack_+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30207 tokenizer_atari_pong_dev_lr_3+         standard  16       FAILED   00:00:01 1-00:00:00 \r\n          30208 tokenizer_atari_alien_dev_lr_+         standard  16       FAILED   00:00:01 1-00:00:00 \r\n          30209 tokenizer_atari_amidar_dev_lr+         standard  16       FAILED   00:00:07 1-00:00:00 \r\n          30210 tokenizer_atari_assault_dev_l+         standard  16       FAILED   00:00:01 1-00:00:00 \r\n          30211 tokenizer_atari_asterix_dev_l+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30212 tokenizer_atari_bank_heist_de+         standard  16       FAILED   00:00:01 1-00:00:00 \r\n          30213 tokenizer_atari_battle_zone_d+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30214 tokenizer_atari_boxing_dev_lr+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30215 tokenizer_atari_breakout_dev_+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30216 tokenizer_atari_chopper_comma+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30217 tokenizer_atari_crazy_climber+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30218 tokenizer_atari_demon_attack_+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30219 tokenizer_atari_pong_dev_lr_3+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30220 tokenizer_atari_alien_dev_lr_+         standard  16       FAILED   00:00:01 1-00:00:00 \r\n          30221 tokenizer_atari_amidar_dev_lr+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30222 tokenizer_atari_assault_dev_l+         standard  16       FAILED   00:00:01 1-00:00:00 \r\n          30223 tokenizer_atari_asterix_dev_l+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30224 tokenizer_atari_bank_heist_de+         standard  16       FAILED   00:00:01 1-00:00:00 \r\n          30225 tokenizer_atari_battle_zone_d+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30226 tokenizer_atari_boxing_dev_lr+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30227 tokenizer_atari_breakout_dev_+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30228 tokenizer_atari_chopper_comma+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30229 tokenizer_atari_crazy_climber+         standard  16       FAILED   00:00:07 1-00:00:00 \r\n          30230 tokenizer_atari_demon_attack_+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30231 tokenizer_atari_pong_dev_lr_3+         standard  16       FAILED   00:00:02 1-00:00:00 \r\n          30245 tokenizer_atari_amidar_dev_lr+         standard  16    COMPLETED   08:52:50 1-00:00:00 \r\n          30247 tokenizer_atari_asterix_dev_l+         standard  16    COMPLETED   09:38:21 1-00:00:00 \r\n          30253 tokenizer_atari_crazy_climber+         standard  16    COMPLETED   20:17:14 1-00:00:00 \r\n          30254 tokenizer_atari_demon_attack_+         standard  16    COMPLETED   17:18:50 1-00:00:00 \r\n          30256 tokenizer_atari_alien_dev_lr_+         standard  16    COMPLETED   07:20:19 1-00:00:00 \r\n          30257 tokenizer_atari_amidar_dev_lr+         standard  16    COMPLETED   04:59:47 1-00:00:00 \r\n          30292 dynamics_atari_alien_dev_150k+         standard  16    COMPLETED   06:12:28 1-00:00:00 \r\n          30293 dynamics_atari_amidar_dev_150+         standard  16    COMPLETED   06:13:27 1-00:00:00 \r\n          30294 dynamics_atari_assault_dev_15+         standard  16    COMPLETED   06:14:02 1-00:00:00 \r\n          30295 dynamics_atari_asterix_dev_15+         standard  16    COMPLETED   06:14:34 1-00:00:00 \r\n          30296 dynamics_atari_bank_heist_dev+         standard  16    COMPLETED   06:16:00 1-00:00:00 \r\n          30297 dynamics_atari_battle_zone_de+         standard  16    COMPLETED   05:15:35 1-00:00:00 \r\n          30298 dynamics_atari_boxing_dev_150+         standard  16    COMPLETED   06:12:18 1-00:00:00 \r\n          30299 dynamics_atari_breakout_dev_1+         standard  16    COMPLETED   06:08:35 1-00:00:00 \r\n          30300 dynamics_atari_chopper_comman+         standard  16    COMPLETED   06:07:29 1-00:00:00 \r\n          30301 dynamics_atari_crazy_climber_+         standard  16    COMPLETED   06:09:03 1-00:00:00 \r\n          30302 dynamics_atari_demon_attack_d+         standard  16    COMPLETED   05:38:48 1-00:00:00 \r\n          30303 dynamics_atari_pong_dev_150k_+         standard  16    COMPLETED   05:38:18 1-00:00:00 \r\n          30280 tokenizer_atari_alien_dev_lr_+         standard  16    COMPLETED   04:22:42 1-00:00:00 \r\n          30281 tokenizer_atari_amidar_dev_lr+         standard  16    COMPLETED   19:11:52 1-00:00:00 \r\n          30284 tokenizer_atari_bank_heist_de+         standard  16    COMPLETED   05:08:15 1-00:00:00 \r\n          30285 tokenizer_atari_battle_zone_d+         standard  16    COMPLETED   19:19:43 1-00:00:00 \r\n          30286 tokenizer_atari_boxing_dev_lr+         standard  16    COMPLETED   12:04:22 1-00:00:00 \r\n          30288 tokenizer_atari_chopper_comma+         standard  16    COMPLETED   10:21:46 1-00:00:00 \r\n          30289 tokenizer_atari_crazy_climber+         standard  16    COMPLETED   22:11:51 1-00:00:00 \r\n          30291 tokenizer_atari_pong_dev_lr_3+         standard  16    COMPLETED   00:40:49 1-00:00:00 \r\n          30282 tokenizer_atari_assault_dev_l+         standard  16    COMPLETED   08:51:24 1-00:00:00 \r\n          30283 tokenizer_atari_asterix_dev_l+         standard  16    COMPLETED   03:38:40 1-00:00:00 \r\n          30287 tokenizer_atari_breakout_dev_+         standard  16    COMPLETED   04:43:13 1-00:00:00 \r\n          30496 dynamics_atari_alien_dev_fleu+         standard  16       FAILED   00:00:49 1-00:00:00 \r\n          30497 dynamics_atari_amidar_dev_fle+         standard  16       FAILED   00:01:11 1-00:00:00 \r\n          30498 dynamics_atari_assault_dev_fl+         standard  16       FAILED   00:00:46 1-00:00:00 \r\n          30499 dynamics_atari_asterix_dev_fl+         standard  16       FAILED   00:00:34 1-00:00:00 \r\n          30500 dynamics_atari_bank_heist_dev+         standard  16       FAILED   00:01:14 1-00:00:00 \r\n          30501 dynamics_atari_battle_zone_de+         standard  16       FAILED   00:00:51 1-00:00:00 \r\n          30502 dynamics_atari_boxing_dev_fle+         standard  16       FAILED   00:00:47 1-00:00:00 \r\n          30503 dynamics_atari_breakout_dev_f+         standard  16       FAILED   00:01:15 1-00:00:00 \r\n          30504 dynamics_atari_crazy_climber_+         standard  16       FAILED   00:00:51 1-00:00:00 \r\n          30505 dynamics_atari_alien_dev_fleu+         standard  16       FAILED   00:00:48 1-00:00:00 \r\n          30506 dynamics_atari_amidar_dev_fle+         standard  16       FAILED   00:00:47 1-00:00:00 \r\n          30507 dynamics_atari_assault_dev_fl+         standard  16       FAILED   00:00:32 1-00:00:00 \r\n          30508 dynamics_atari_asterix_dev_fl+         standard  16       FAILED   00:00:48 1-00:00:00 \r\n          30509 dynamics_atari_bank_heist_dev+         standard  16       FAILED   00:00:51 1-00:00:00 \r\n          30510 dynamics_atari_battle_zone_de+         standard  16       FAILED   00:00:34 1-00:00:00 \r\n          30511 dynamics_atari_boxing_dev_fle+         standard  16       FAILED   00:00:32 1-00:00:00 \r\n          30512 dynamics_atari_breakout_dev_f+         standard  16       FAILED   00:01:02 1-00:00:00 \r\n          30513 dynamics_atari_crazy_climber_+         standard  16       FAILED   00:00:31 1-00:00:00 \r\n          30514 dynamics_atari_alien_dev_fleu+         standard  16       FAILED   00:00:38 1-00:00:00 \r\n          30515 dynamics_atari_amidar_dev_fle+         standard  16       FAILED   00:00:46 1-00:00:00 \r\n          30516 dynamics_atari_assault_dev_fl+         standard  16       FAILED   00:00:48 1-00:00:00 \r\n          30517 dynamics_atari_asterix_dev_fl+         standard  16       FAILED   00:00:47 1-00:00:00 \r\n          30518 dynamics_atari_bank_heist_dev+         standard  16       FAILED   00:00:36 1-00:00:00 \r\n          30519 dynamics_atari_battle_zone_de+         standard  16       FAILED   00:00:38 1-00:00:00 \r\n          30520 dynamics_atari_boxing_dev_fle+         standard  16       FAILED   00:01:05 1-00:00:00 \r\n          30521 dynamics_atari_breakout_dev_f+         standard  16       FAILED   00:00:36 1-00:00:00 \r\n          30522 dynamics_atari_crazy_climber_+         standard  16       FAILED   00:00:46 1-00:00:00 \r\n          30523 dynamics_atari_alien_dev_fleu+         standard  16    COMPLETED   20:42:47 1-00:00:00 \r\n          30524 dynamics_atari_amidar_dev_fle+         standard  16    COMPLETED   20:42:23 1-00:00:00 \r\n          30525 dynamics_atari_assault_dev_fl+         standard  16    COMPLETED   20:47:41 1-00:00:00 \r\n          30526 dynamics_atari_asterix_dev_fl+         standard  16    COMPLETED   20:42:56 1-00:00:00 \r\n          30527 dynamics_atari_bank_heist_dev+         standard  16    COMPLETED   20:43:41 1-00:00:00 \r\n          30528 dynamics_atari_battle_zone_de+         standard  16    COMPLETED   20:46:16 1-00:00:00 \r\n          30529 dynamics_atari_boxing_dev_fle+         standard  16    COMPLETED   20:42:42 1-00:00:00 \r\n          30530 dynamics_atari_breakout_dev_f+         standard  16    COMPLETED   20:47:02 1-00:00:00 \r\n          30531 dynamics_atari_crazy_climber_+         standard  16    COMPLETED   20:44:45 1-00:00:00 \r\n          30290 tokenizer_atari_demon_attack_+         standard  16    COMPLETED   02:19:41 1-00:00:00 \r\n          31727 dynamics_atari_demon_attack_d+         standard  16    COMPLETED   20:47:08 1-00:00:00 \r\n          31761     sample_atari_alien_maskgit         standard  16    COMPLETED   00:01:13 1-00:00:00 \r\n          31762    sample_atari_amidar_maskgit         standard  16    COMPLETED   00:01:17 1-00:00:00 \r\n          31763   sample_atari_assault_maskgit         standard  16    COMPLETED   00:01:13 1-00:00:00 \r\n          31764   sample_atari_asterix_maskgit         standard  16    COMPLETED   00:01:12 1-00:00:00 \r\n          31765 sample_atari_bank_heist_maskg+         standard  16    COMPLETED   00:01:23 1-00:00:00 \r\n          31766 sample_atari_battle_zone_mask+         standard  16    COMPLETED   00:01:06 1-00:00:00 \r\n          31767  sample_atari_breakout_maskgit         standard  16    COMPLETED   00:01:01 1-00:00:00 \r\n          31768 sample_atari_crazy_climber_ma+         standard  16    COMPLETED   00:01:03 1-00:00:00 \r\n          31863 coinrun_sample_maskgit_mila_s+         standard  16       FAILED   00:00:14 1-00:00:00 \r\n          31864 coinrun_sample_maskgit_mila_s+         standard  16       FAILED   00:00:19 1-00:00:00 \r\n          31865 coinrun_sample_maskgit_mila_s+         standard  16       FAILED   00:00:13 1-00:00:00 \r\n          31866 coinrun_sample_maskgit_mila_s+         standard  16       FAILED   00:00:12 1-00:00:00 \r\n          31867 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:00 1-00:00:00 \r\n          31868 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:00:57 1-00:00:00 \r\n          31869 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:02 1-00:00:00 \r\n          31870 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:25 1-00:00:00 \r\n          31871 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:26 1-00:00:00 \r\n          31872 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:21 1-00:00:00 \r\n          31873 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:25 1-00:00:00 \r\n          31874 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:21 1-00:00:00 \r\n          31875 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:24 1-00:00:00 \r\n          31882 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:28 1-00:00:00 \r\n          31883 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:29 1-00:00:00 \r\n          32059 coinrun_sample_maskgit_mila_s+         standard  16       FAILED   00:00:01 1-00:00:00 \r\n          32060 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:11 1-00:00:00 \r\n          32061 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:00:57 1-00:00:00 \r\n          32062 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:03 1-00:00:00 \r\n          32063 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:01 1-00:00:00 \r\n]0;franz.srambical@hai-login2:~/jafar",,terminal_output
+13,117951,"TERMINAL",0,0,"",,terminal_command
+14,123411,"TERMINAL",0,0,"sacct --format=""JobID%15,JobName%30,Partition%16,AllocCPUS%3,State%12,Elapsed%10,Timelimit%10,Start,End"" --starttime=$(date -d ""last month"" +%Y-%m-%d) | grep -vE ""*.batch|*.extern|*.inter|bash|python|CANCELLED|echo""",,terminal_command
+15,123471,"TERMINAL",0,0,"]633;C          JobID                        JobName        Partition All        State    Elapsed  Timelimit               Start                 End \r\n--------------- ------------------------------ ---------------- --- ------------ ---------- ---------- ------------------- ------------------- \r\n          29284                           wrap         standard 128    COMPLETED   10:09:39 1-00:00:00 2025-09-22T13:58:28 2025-09-23T00:08:07 \r\n          29399            generate_atari_data         standard  16    COMPLETED   01:03:04 1-00:00:00 2025-09-24T18:52:28 2025-09-24T19:55:32 \r\n          29462            generate_atari_data         standard 128    COMPLETED   15:12:02 1-00:00:00 2025-09-25T14:55:44 2025-09-26T06:07:46 \r\n          29473 generate_minatar_breakout_data         standard 128    COMPLETED   02:02:11 1-00:00:00 2025-09-25T16:01:20 2025-09-25T18:03:31 \r\n          29671 generate_atari_breakout_no_st+         standard 128    COMPLETED   08:39:39 1-00:00:00 2025-09-29T22:50:22 2025-09-30T07:30:01 \r\n          29672 generate_atari_pong_no_sticky+         standard 128    COMPLETED   08:41:40 1-00:00:00 2025-09-29T22:50:37 2025-09-30T07:32:17 \r\n          29673 generate_atari_alien_no_stick+         standard 128       FAILED   00:00:14 1-00:00:00 2025-09-29T23:29:04 2025-09-29T23:29:18 \r\n          29674 generate_atari_amidar_no_stic+         standard 128       FAILED   00:00:14 1-00:00:00 2025-09-29T23:29:04 2025-09-29T23:29:18 \r\n          29675 generate_atari_assault_no_sti+         standard 128       FAILED   00:00:20 1-00:00:00 2025-09-29T23:29:04 2025-09-29T23:29:24 \r\n          29676 generate_atari_asterix_no_sti+         standard 128       FAILED   00:00:10 1-00:00:00 2025-09-29T23:29:18 2025-09-29T23:29:28 \r\n          29677 generate_atari_bank_heist_no_+         standard 128       FAILED   00:00:10 1-00:00:00 2025-09-29T23:29:23 2025-09-29T23:29:33 \r\n          29678 generate_atari_battle_zone_no+         standard 128       FAILED   00:00:10 1-00:00:00 2025-09-29T23:29:28 2025-09-29T23:29:38 \r\n          29679 generate_atari_boxing_no_stic+         standard 128       FAILED   00:00:10 1-00:00:00 2025-09-29T23:29:33 2025-09-29T23:29:43 \r\n          29680 generate_atari_chopper_comman+         standard 128       FAILED   00:00:10 1-00:00:00 2025-09-29T23:29:34 2025-09-29T23:29:44 \r\n          29681 generate_atari_crazy_climber_+         standard 128       FAILED   00:00:10 1-00:00:00 2025-09-29T23:29:38 2025-09-29T23:29:48 \r\n          29682 generate_atari_demon_attack_n+         standard 128       FAILED   00:00:09 1-00:00:00 2025-09-29T23:29:43 2025-09-29T23:29:52 \r\n          29683 generate_atari_alien_no_stick+         standard 128    COMPLETED   09:46:52 1-00:00:00 2025-09-29T23:32:46 2025-09-30T09:19:38 \r\n          29685 generate_atari_amidar_no_stic+         standard 128    COMPLETED   09:21:44 1-00:00:00 2025-09-29T23:35:59 2025-09-30T08:57:43 \r\n          29686 generate_atari_assault_no_sti+         standard 128    COMPLETED   08:32:12 1-00:00:00 2025-09-29T23:35:59 2025-09-30T08:08:11 \r\n          29687 generate_atari_asterix_no_sti+         standard 128    COMPLETED   08:51:17 1-00:00:00 2025-09-29T23:44:15 2025-09-30T08:35:32 \r\n          29688 generate_atari_bank_heist_no_+         standard 128    COMPLETED   09:02:48 1-00:00:00 2025-09-30T07:30:01 2025-09-30T16:32:49 \r\n          29689 generate_atari_battle_zone_no+         standard 128    COMPLETED   09:20:54 1-00:00:00 2025-09-30T07:32:17 2025-09-30T16:53:11 \r\n          29690 generate_atari_boxing_no_stic+         standard 128    COMPLETED   09:18:27 1-00:00:00 2025-09-30T08:08:11 2025-09-30T17:26:38 \r\n          29691 generate_atari_chopper_comman+         standard 128    COMPLETED   09:17:26 1-00:00:00 2025-09-30T08:35:32 2025-09-30T17:52:58 \r\n          29692 generate_atari_crazy_climber_+         standard 128    COMPLETED   08:49:47 1-00:00:00 2025-09-30T08:57:44 2025-09-30T17:47:31 \r\n          29693 generate_atari_demon_attack_n+         standard 128    COMPLETED   08:16:52 1-00:00:00 2025-09-30T09:19:38 2025-09-30T17:36:30 \r\n          29734 tokenizer_coinrun_mila_submis+         standard  16       FAILED   00:00:00 1-00:00:00 2025-09-30T18:48:10 2025-09-30T18:48:10 \r\n          29736 tokenizer_coinrun_mila_submis+         standard  16    COMPLETED   03:58:28 1-00:00:00 2025-09-30T18:52:14 2025-09-30T22:50:42 \r\n          29737    lam_coinrun_mila_submission         standard  16    COMPLETED   02:31:23 1-00:00:00 2025-09-30T19:02:41 2025-09-30T21:34:04 \r\n          29738 lam_coinrun_mila_submission_n+         standard  16    COMPLETED   02:35:49 1-00:00:00 2025-09-30T19:05:04 2025-09-30T21:40:53 \r\n          29759 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   06:46:34 1-00:00:00 2025-09-30T23:15:31 2025-10-01T06:02:05 \r\n          29760 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   04:02:30 1-00:00:00 2025-09-30T23:16:01 2025-10-01T03:18:31 \r\n          29761 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   05:09:04 1-00:00:00 2025-09-30T23:16:01 2025-10-01T04:25:05 \r\n          29762 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   05:22:21 1-00:00:00 2025-09-30T23:16:31 2025-10-01T04:38:52 \r\n          29763 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   05:11:06 1-00:00:00 2025-09-30T23:16:31 2025-10-01T04:27:37 \r\n          29764 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:21 1-00:00:00 2025-09-30T23:22:02 2025-09-30T23:22:23 \r\n          29765 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:21 1-00:00:00 2025-09-30T23:28:32 2025-09-30T23:28:53 \r\n          29766 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   05:35:34 1-00:00:00 2025-09-30T23:30:32 2025-10-01T05:06:06 \r\n          29771 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   14:03:09 1-00:00:00 2025-10-01T09:47:42 2025-10-01T23:50:51 \r\n          29772 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   14:47:34 1-00:00:00 2025-10-01T09:51:53 2025-10-02T00:39:27 \r\n          29773 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   04:36:48 1-00:00:00 2025-10-01T09:52:29 2025-10-01T14:29:17 \r\n          29804 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:01 1-00:00:00 2025-10-01T16:04:13 2025-10-01T16:04:14 \r\n          29805 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:00 1-00:00:00 2025-10-01T16:04:33 2025-10-01T16:04:33 \r\n          29807 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   14:44:21 1-00:00:00 2025-10-01T16:08:28 2025-10-02T06:52:49 \r\n          29808 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   04:29:03 1-00:00:00 2025-10-01T16:08:48 2025-10-01T20:37:51 \r\n          29845 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   04:24:16 1-00:00:00 2025-10-01T19:14:47 2025-10-01T23:39:03 \r\n          29898 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:01 1-00:00:00 2025-10-02T12:16:40 2025-10-02T12:16:41 \r\n          29899 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   03:25:34 1-00:00:00 2025-10-02T12:18:34 2025-10-02T15:44:08 \r\n          29906 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:01:16 1-00:00:00 2025-10-02T18:22:26 2025-10-02T18:23:42 \r\n          29923 tokenizer_coinrun_mila_submis+         standard  16       FAILED   02:22:42 1-00:00:00 2025-10-02T19:43:56 2025-10-02T22:06:38 \r\n          29924    lam_coinrun_mila_submission         standard  16       FAILED   02:21:54 1-00:00:00 2025-10-02T19:45:26 2025-10-02T22:07:20 \r\n          29927 tokenizer_coinrun_mila_submis+         standard  16       FAILED   01:04:49 1-00:00:00 2025-10-02T23:17:28 2025-10-03T00:22:17 \r\n          29928    lam_coinrun_mila_submission         standard  16       FAILED   01:04:47 1-00:00:00 2025-10-02T23:17:28 2025-10-03T00:22:15 \r\n          29949 tokenizer_coinrun_mila_submis+         standard  16    COMPLETED   03:48:24 1-00:00:00 2025-10-03T13:05:25 2025-10-03T16:53:49 \r\n          29950    lam_coinrun_mila_submission         standard  16    COMPLETED   02:20:48 1-00:00:00 2025-10-03T13:05:45 2025-10-03T15:26:33 \r\n          29959 lam_coinrun_mila_submission_n+         standard  16    COMPLETED   02:24:53 1-00:00:00 2025-10-03T14:24:26 2025-10-03T16:49:19 \r\n          29973 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   05:01:33 1-00:00:00 2025-10-03T17:12:22 2025-10-03T22:13:55 \r\n          29974 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   04:59:19 1-00:00:00 2025-10-03T17:13:45 2025-10-03T22:13:04 \r\n          29976 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   06:33:08 1-00:00:00 2025-10-03T17:14:28 2025-10-03T23:47:36 \r\n          29977 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   03:45:18 1-00:00:00 2025-10-03T17:15:27 2025-10-03T21:00:45 \r\n          29978 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   05:03:54 1-00:00:00 2025-10-03T17:16:48 2025-10-03T22:20:42 \r\n          29985 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   02:28:50 1-00:00:00 2025-10-03T18:02:27 2025-10-03T20:31:17 \r\n          29986 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   03:13:31 1-00:00:00 2025-10-03T18:02:54 2025-10-03T21:16:25 \r\n          29987 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   02:36:47 1-00:00:00 2025-10-03T18:03:40 2025-10-03T20:40:27 \r\n          29988 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:01 1-00:00:00 2025-10-03T18:10:33 2025-10-03T18:10:34 \r\n          29989 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:01 1-00:00:00 2025-10-03T18:12:42 2025-10-03T18:12:43 \r\n          29990 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   02:34:01 1-00:00:00 2025-10-03T18:15:54 2025-10-03T20:49:55 \r\n          29991 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:01 1-00:00:00 2025-10-03T18:28:20 2025-10-03T18:28:21 \r\n          29995 dynamics_coinrun_mila_submiss+         standard  16       FAILED   00:00:45 1-00:00:00 2025-10-03T18:33:19 2025-10-03T18:34:04 \r\n          29996 dynamics_coinrun_mila_submiss+         standard  16    COMPLETED   19:46:35 1-00:00:00 2025-10-03T18:38:38 2025-10-04T14:25:13 \r\n          30047      tokenizer_atari_alien_dev         standard  16       FAILED   00:00:00 1-00:00:00 2025-10-04T12:34:26 2025-10-04T12:34:26 \r\n          30048      tokenizer_atari_alien_dev         standard  16       FAILED   00:00:32 1-00:00:00 2025-10-04T12:35:26 2025-10-04T12:35:58 \r\n          30075           lam_atari_amidar_dev         standard  16    COMPLETED   03:53:54 1-00:00:00 2025-10-04T14:25:26 2025-10-04T18:19:20 \r\n          30076          lam_atari_assault_dev         standard  16    COMPLETED   03:52:52 1-00:00:00 2025-10-04T14:49:26 2025-10-04T18:42:18 \r\n          30077          lam_atari_asterix_dev         standard  16    COMPLETED   03:48:53 1-00:00:00 2025-10-04T14:49:26 2025-10-04T18:38:19 \r\n          30078       lam_atari_bank_heist_dev         standard  16    COMPLETED   02:58:29 1-00:00:00 2025-10-04T14:49:26 2025-10-04T17:47:55 \r\n          30079      lam_atari_battle_zone_dev         standard  16    COMPLETED   03:54:36 1-00:00:00 2025-10-04T14:49:26 2025-10-04T18:44:02 \r\n          30080           lam_atari_boxing_dev         standard  16    COMPLETED   02:59:00 1-00:00:00 2025-10-04T14:49:26 2025-10-04T17:48:26 \r\n          30081         lam_atari_breakout_dev         standard  16    COMPLETED   04:05:11 1-00:00:00 2025-10-04T14:50:26 2025-10-04T18:55:37 \r\n          30082  lam_atari_chopper_command_dev         standard  16       FAILED   00:00:20 1-00:00:00 2025-10-04T14:50:26 2025-10-04T14:50:46 \r\n          30083    lam_atari_crazy_climber_dev         standard  16    COMPLETED   04:08:20 1-00:00:00 2025-10-04T14:50:26 2025-10-04T18:58:46 \r\n          30084     lam_atari_demon_attack_dev         standard  16    COMPLETED   04:09:50 1-00:00:00 2025-10-04T14:50:26 2025-10-04T19:00:16 \r\n          30085             lam_atari_pong_dev         standard  16    COMPLETED   04:09:13 1-00:00:00 2025-10-04T14:50:26 2025-10-04T18:59:39 \r\n          30092      tokenizer_atari_alien_dev         standard  16    COMPLETED   02:17:47 1-00:00:00 2025-10-04T14:55:26 2025-10-04T17:13:13 \r\n          30093     tokenizer_atari_amidar_dev         standard  16    COMPLETED   01:48:07 1-00:00:00 2025-10-04T14:55:26 2025-10-04T16:43:33 \r\n          30094    tokenizer_atari_assault_dev         standard  16    COMPLETED   03:05:43 1-00:00:00 2025-10-04T14:55:26 2025-10-04T18:01:09 \r\n          30095    tokenizer_atari_asterix_dev         standard  16    COMPLETED   02:03:04 1-00:00:00 2025-10-04T15:01:56 2025-10-04T17:05:00 \r\n          30096 tokenizer_atari_bank_heist_dev         standard  16    COMPLETED   01:27:56 1-00:00:00 2025-10-04T16:35:56 2025-10-04T18:03:52 \r\n          30097 tokenizer_atari_battle_zone_d+         standard  16    COMPLETED   01:13:24 1-00:00:00 2025-10-04T16:43:56 2025-10-04T17:57:20 \r\n          30098     tokenizer_atari_boxing_dev         standard  16    COMPLETED   01:37:54 1-00:00:00 2025-10-04T17:05:27 2025-10-04T18:43:21 \r\n          30099   tokenizer_atari_breakout_dev         standard  16    COMPLETED   01:49:10 1-00:00:00 2025-10-04T17:13:27 2025-10-04T19:02:37 \r\n          30100 tokenizer_atari_chopper_comma+         standard  16    COMPLETED   01:16:14 1-00:00:00 2025-10-04T17:45:57 2025-10-04T19:02:11 \r\n          30101 tokenizer_atari_crazy_climber+         standard  16    COMPLETED   03:19:30 1-00:00:00 2025-10-04T17:47:57 2025-10-04T21:07:27 \r\n          30102 tokenizer_atari_demon_attack_+         standard  16    COMPLETED   01:57:06 1-00:00:00 2025-10-04T17:48:27 2025-10-04T19:45:33 \r\n          30103       tokenizer_atari_pong_dev         standard  16    COMPLETED   01:30:45 1-00:00:00 2025-10-04T17:52:27 2025-10-04T19:23:12 \r\n          30117       dynamics_atari_alien_dev         standard  16    COMPLETED   06:10:49 1-00:00:00 2025-10-04T18:35:27 2025-10-05T00:46:16 \r\n          30118      dynamics_atari_amidar_dev         standard  16    COMPLETED   05:23:49 1-00:00:00 2025-10-04T18:35:27 2025-10-04T23:59:16 \r\n          30119     dynamics_atari_assault_dev         standard  16    COMPLETED   06:20:45 1-00:00:00 2025-10-04T18:35:27 2025-10-05T00:56:12 \r\n          30120     dynamics_atari_asterix_dev         standard  16    COMPLETED   05:27:53 1-00:00:00 2025-10-04T18:35:27 2025-10-05T00:03:20 \r\n          30121  dynamics_atari_bank_heist_dev         standard  16    COMPLETED   05:28:06 1-00:00:00 2025-10-04T18:35:27 2025-10-05T00:03:33 \r\n          30122 dynamics_atari_battle_zone_dev         standard  16    COMPLETED   06:11:44 1-00:00:00 2025-10-04T18:38:27 2025-10-05T00:50:11 \r\n          30123      dynamics_atari_boxing_dev         standard  16    COMPLETED   06:14:07 1-00:00:00 2025-10-04T18:42:27 2025-10-05T00:56:34 \r\n          30124    dynamics_atari_breakout_dev         standard  16    COMPLETED   06:09:47 1-00:00:00 2025-10-04T18:43:27 2025-10-05T00:53:14 \r\n          30125 dynamics_atari_chopper_comman+         standard  16    COMPLETED   06:12:46 1-00:00:00 2025-10-04T18:44:27 2025-10-05T00:57:13 \r\n          30126        dynamics_atari_pong_dev         standard  16    COMPLETED   06:19:33 1-00:00:00 2025-10-04T18:55:57 2025-10-05T01:15:30 \r\n          30127 dynamics_atari_demon_attack_d+         standard  16    COMPLETED   06:16:14 1-00:00:00 2025-10-04T18:58:57 2025-10-05T01:15:11 \r\n          30130 tokenizer_atari_alien_dev_lr_+         standard  16    COMPLETED   03:07:39 1-00:00:00 2025-10-04T19:18:33 2025-10-04T22:26:12 \r\n          30131 tokenizer_atari_amidar_dev_lr+         standard  16    COMPLETED   02:13:25 1-00:00:00 2025-10-04T19:18:34 2025-10-04T21:31:59 \r\n          30132 tokenizer_atari_assault_dev_l+         standard  16    COMPLETED   03:28:04 1-00:00:00 2025-10-04T19:18:34 2025-10-04T22:46:38 \r\n          30133 tokenizer_atari_asterix_dev_l+         standard  16    COMPLETED   04:04:16 1-00:00:00 2025-10-04T19:18:34 2025-10-04T23:22:50 \r\n          30134 tokenizer_atari_bank_heist_de+         standard  16    COMPLETED   01:33:50 1-00:00:00 2025-10-04T19:18:34 2025-10-04T20:52:24 \r\n          30135 tokenizer_atari_battle_zone_d+         standard  16    COMPLETED   01:44:19 1-00:00:00 2025-10-04T19:23:13 2025-10-04T21:07:32 \r\n          30136 tokenizer_atari_boxing_dev_lr+         standard  16    COMPLETED   01:58:10 1-00:00:00 2025-10-04T19:45:35 2025-10-04T21:43:45 \r\n          30137 tokenizer_atari_breakout_dev_+         standard  16    COMPLETED   01:52:02 1-00:00:00 2025-10-04T20:52:40 2025-10-04T22:44:42 \r\n          30138 tokenizer_atari_chopper_comma+         standard  16    COMPLETED   01:28:25 1-00:00:00 2025-10-04T21:07:40 2025-10-04T22:36:05 \r\n          30139 tokenizer_atari_crazy_climber+         standard  16    COMPLETED   04:37:01 1-00:00:00 2025-10-04T21:07:40 2025-10-05T01:44:41 \r\n          30140 tokenizer_atari_demon_attack_+         standard  16    COMPLETED   03:58:56 1-00:00:00 2025-10-04T21:32:10 2025-10-05T01:31:06 \r\n          30141 tokenizer_atari_pong_dev_lr_3+         standard  16    COMPLETED   01:46:00 1-00:00:00 2025-10-04T21:44:10 2025-10-04T23:30:10 \r\n          30143 tokenizer_atari_alien_dev_lr_+         standard  16       FAILED   00:00:25 1-00:00:00 2025-10-04T22:26:40 2025-10-04T22:27:05 \r\n          30144 tokenizer_atari_amidar_dev_lr+         standard  16       FAILED   00:00:32 1-00:00:00 2025-10-04T22:27:10 2025-10-04T22:27:42 \r\n          30145 tokenizer_atari_assault_dev_l+         standard  16       FAILED   00:00:33 1-00:00:00 2025-10-04T22:28:10 2025-10-04T22:28:43 \r\n          30146 tokenizer_atari_asterix_dev_l+         standard  16       FAILED   00:00:34 1-00:00:00 2025-10-04T22:29:10 2025-10-04T22:29:44 \r\n          30147 tokenizer_atari_bank_heist_de+         standard  16       FAILED   00:00:41 1-00:00:00 2025-10-04T22:30:10 2025-10-04T22:30:51 \r\n          30148 tokenizer_atari_battle_zone_d+         standard  16       FAILED   00:00:38 1-00:00:00 2025-10-04T22:31:10 2025-10-04T22:31:48 \r\n          30149 tokenizer_atari_boxing_dev_lr+         standard  16       FAILED   00:00:30 1-00:00:00 2025-10-04T22:32:10 2025-10-04T22:32:40 \r\n          30150 tokenizer_atari_breakout_dev_+         standard  16       FAILED   00:00:29 1-00:00:00 2025-10-04T22:32:40 2025-10-04T22:33:09 \r\n          30151 tokenizer_atari_chopper_comma+         standard  16       FAILED   00:00:31 1-00:00:00 2025-10-04T22:33:10 2025-10-04T22:33:41 \r\n          30152 tokenizer_atari_crazy_climber+         standard  16       FAILED   00:00:34 1-00:00:00 2025-10-04T22:34:10 2025-10-04T22:34:44 \r\n          30153 tokenizer_atari_demon_attack_+         standard  16       FAILED   00:00:37 1-00:00:00 2025-10-04T22:35:10 2025-10-04T22:35:47 \r\n          30154 tokenizer_atari_pong_dev_lr_3+         standard  16       FAILED   00:00:32 1-00:00:00 2025-10-04T22:36:10 2025-10-04T22:36:42 \r\n          30155 tokenizer_atari_alien_dev_lr_+         standard  16       FAILED   00:00:43 1-00:00:00 2025-10-04T22:36:10 2025-10-04T22:36:53 \r\n          30156 tokenizer_atari_amidar_dev_lr+         standard  16       FAILED   00:00:36 1-00:00:00 2025-10-04T22:37:10 2025-10-04T22:37:46 \r\n          30157 tokenizer_atari_assault_dev_l+         standard  16       FAILED   00:00:28 1-00:00:00 2025-10-04T22:37:10 2025-10-04T22:37:38 \r\n          30158 tokenizer_atari_asterix_dev_l+         standard  16       FAILED   00:00:37 1-00:00:00 2025-10-04T22:37:40 2025-10-04T22:38:17 \r\n          30159 tokenizer_atari_bank_heist_de+         standard  16       FAILED   00:00:36 1-00:00:00 2025-10-04T22:38:10 2025-10-04T22:38:46 \r\n          30160 tokenizer_atari_battle_zone_d+         standard  16       FAILED   00:00:32 1-00:00:00 2025-10-04T22:38:40 2025-10-04T22:39:12 \r\n          30161 tokenizer_atari_boxing_dev_lr+         standard  16       FAILED   00:00:30 1-00:00:00 2025-10-04T22:39:10 2025-10-04T22:39:40 \r\n          30162 tokenizer_atari_breakout_dev_+         standard  16       FAILED   00:00:41 1-00:00:00 2025-10-04T22:39:40 2025-10-04T22:40:21 \r\n          30163 tokenizer_atari_chopper_comma+         standard  16       FAILED   00:00:21 1-00:00:00 2025-10-04T22:40:10 2025-10-04T22:40:31 \r\n          30164 tokenizer_atari_crazy_climber+         standard  16       FAILED   00:00:28 1-00:00:00 2025-10-04T22:40:40 2025-10-04T22:41:08 \r\n          30165 tokenizer_atari_demon_attack_+         standard  16       FAILED   00:00:38 1-00:00:00 2025-10-04T22:40:40 2025-10-04T22:41:18 \r\n          30166 tokenizer_atari_pong_dev_lr_3+         standard  16       FAILED   00:00:27 1-00:00:00 2025-10-04T22:41:10 2025-10-04T22:41:37 \r\n          30167 tokenizer_atari_alien_dev_lr_+         standard  16       FAILED   00:00:29 1-00:00:00 2025-10-04T22:41:40 2025-10-04T22:42:09 \r\n          30168 tokenizer_atari_amidar_dev_lr+         standard  16       FAILED   00:00:42 1-00:00:00 2025-10-04T22:41:40 2025-10-04T22:42:22 \r\n          30169 tokenizer_atari_assault_dev_l+         standard  16       FAILED   00:00:37 1-00:00:00 2025-10-04T22:42:10 2025-10-04T22:42:47 \r\n          30170 tokenizer_atari_asterix_dev_l+         standard  16       FAILED   00:00:33 1-00:00:00 2025-10-04T22:42:40 2025-10-04T22:43:13 \r\n          30171 tokenizer_atari_bank_heist_de+         standard  16       FAILED   00:00:35 1-00:00:00 2025-10-04T22:43:10 2025-10-04T22:43:45 \r\n          30172 tokenizer_atari_battle_zone_d+         standard  16       FAILED   00:00:38 1-00:00:00 2025-10-04T22:43:40 2025-10-04T22:44:18 \r\n          30173 tokenizer_atari_boxing_dev_lr+         standard  16       FAILED   00:00:26 1-00:00:00 2025-10-04T22:44:10 2025-10-04T22:44:36 \r\n          30174 tokenizer_atari_breakout_dev_+         standard  16       FAILED   00:00:30 1-00:00:00 2025-10-04T22:44:40 2025-10-04T22:45:10 \r\n          30175 tokenizer_atari_chopper_comma+         standard  16       FAILED   00:00:37 1-00:00:00 2025-10-04T22:44:40 2025-10-04T22:45:17 \r\n          30176 tokenizer_atari_crazy_climber+         standard  16       FAILED   00:00:32 1-00:00:00 2025-10-04T22:45:10 2025-10-04T22:45:42 \r\n          30177 tokenizer_atari_demon_attack_+         standard  16       FAILED   00:00:43 1-00:00:00 2025-10-04T22:45:10 2025-10-04T22:45:53 \r\n          30178 tokenizer_atari_pong_dev_lr_3+         standard  16       FAILED   00:00:33 1-00:00:00 2025-10-04T22:45:40 2025-10-04T22:46:13 \r\n          30193    doom_dataset_generation_10m         standard  16       FAILED   00:00:06 1-00:00:00 2025-10-05T13:14:56 2025-10-05T13:15:02 \r\n          30194    doom_dataset_generation_10m         standard  16       FAILED   00:00:19 1-00:00:00 2025-10-05T13:16:56 2025-10-05T13:17:15 \r\n          30195    doom_dataset_generation_10m         standard  16       FAILED   00:00:09 1-00:00:00 2025-10-05T13:21:26 2025-10-05T13:21:35 \r\n          30196 tokenizer_atari_alien_dev_lr_+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:45:56 2025-10-05T13:45:58 \r\n          30197 tokenizer_atari_amidar_dev_lr+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:45:56 2025-10-05T13:45:58 \r\n          30198 tokenizer_atari_assault_dev_l+         standard  16       FAILED   00:00:01 1-00:00:00 2025-10-05T13:45:56 2025-10-05T13:45:57 \r\n          30199 tokenizer_atari_asterix_dev_l+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:45:56 2025-10-05T13:45:58 \r\n          30200 tokenizer_atari_bank_heist_de+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:45:56 2025-10-05T13:45:58 \r\n          30201 tokenizer_atari_battle_zone_d+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:45:56 2025-10-05T13:45:58 \r\n          30202 tokenizer_atari_boxing_dev_lr+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:45:56 2025-10-05T13:45:58 \r\n          30203 tokenizer_atari_breakout_dev_+         standard  16       FAILED   00:00:01 1-00:00:00 2025-10-05T13:45:56 2025-10-05T13:45:57 \r\n          30204 tokenizer_atari_chopper_comma+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:45:56 2025-10-05T13:45:58 \r\n          30205 tokenizer_atari_crazy_climber+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:45:56 2025-10-05T13:45:58 \r\n          30206 tokenizer_atari_demon_attack_+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:45:56 2025-10-05T13:45:58 \r\n          30207 tokenizer_atari_pong_dev_lr_3+         standard  16       FAILED   00:00:01 1-00:00:00 2025-10-05T13:45:56 2025-10-05T13:45:57 \r\n          30208 tokenizer_atari_alien_dev_lr_+         standard  16       FAILED   00:00:01 1-00:00:00 2025-10-05T13:47:26 2025-10-05T13:47:27 \r\n          30209 tokenizer_atari_amidar_dev_lr+         standard  16       FAILED   00:00:07 1-00:00:00 2025-10-05T13:47:26 2025-10-05T13:47:33 \r\n          30210 tokenizer_atari_assault_dev_l+         standard  16       FAILED   00:00:01 1-00:00:00 2025-10-05T13:47:26 2025-10-05T13:47:27 \r\n          30211 tokenizer_atari_asterix_dev_l+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:47:26 2025-10-05T13:47:28 \r\n          30212 tokenizer_atari_bank_heist_de+         standard  16       FAILED   00:00:01 1-00:00:00 2025-10-05T13:47:26 2025-10-05T13:47:27 \r\n          30213 tokenizer_atari_battle_zone_d+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:47:26 2025-10-05T13:47:28 \r\n          30214 tokenizer_atari_boxing_dev_lr+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:47:26 2025-10-05T13:47:28 \r\n          30215 tokenizer_atari_breakout_dev_+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:47:26 2025-10-05T13:47:28 \r\n          30216 tokenizer_atari_chopper_comma+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:47:26 2025-10-05T13:47:28 \r\n          30217 tokenizer_atari_crazy_climber+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:47:26 2025-10-05T13:47:28 \r\n          30218 tokenizer_atari_demon_attack_+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:47:26 2025-10-05T13:47:28 \r\n          30219 tokenizer_atari_pong_dev_lr_3+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:47:26 2025-10-05T13:47:28 \r\n          30220 tokenizer_atari_alien_dev_lr_+         standard  16       FAILED   00:00:01 1-00:00:00 2025-10-05T13:56:56 2025-10-05T13:56:57 \r\n          30221 tokenizer_atari_amidar_dev_lr+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:56:56 2025-10-05T13:56:58 \r\n          30222 tokenizer_atari_assault_dev_l+         standard  16       FAILED   00:00:01 1-00:00:00 2025-10-05T13:56:56 2025-10-05T13:56:57 \r\n          30223 tokenizer_atari_asterix_dev_l+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:56:56 2025-10-05T13:56:58 \r\n          30224 tokenizer_atari_bank_heist_de+         standard  16       FAILED   00:00:01 1-00:00:00 2025-10-05T13:56:56 2025-10-05T13:56:57 \r\n          30225 tokenizer_atari_battle_zone_d+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:56:56 2025-10-05T13:56:58 \r\n          30226 tokenizer_atari_boxing_dev_lr+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:56:56 2025-10-05T13:56:58 \r\n          30227 tokenizer_atari_breakout_dev_+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:56:56 2025-10-05T13:56:58 \r\n          30228 tokenizer_atari_chopper_comma+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:56:56 2025-10-05T13:56:58 \r\n          30229 tokenizer_atari_crazy_climber+         standard  16       FAILED   00:00:07 1-00:00:00 2025-10-05T13:56:56 2025-10-05T13:57:03 \r\n          30230 tokenizer_atari_demon_attack_+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:56:56 2025-10-05T13:56:58 \r\n          30231 tokenizer_atari_pong_dev_lr_3+         standard  16       FAILED   00:00:02 1-00:00:00 2025-10-05T13:56:56 2025-10-05T13:56:58 \r\n          30245 tokenizer_atari_amidar_dev_lr+         standard  16    COMPLETED   08:52:50 1-00:00:00 2025-10-05T15:29:56 2025-10-06T00:22:46 \r\n          30247 tokenizer_atari_asterix_dev_l+         standard  16    COMPLETED   09:38:21 1-00:00:00 2025-10-05T15:29:56 2025-10-06T01:08:17 \r\n          30253 tokenizer_atari_crazy_climber+         standard  16    COMPLETED   20:17:14 1-00:00:00 2025-10-05T15:29:56 2025-10-06T11:47:10 \r\n          30254 tokenizer_atari_demon_attack_+         standard  16    COMPLETED   17:18:50 1-00:00:00 2025-10-05T15:29:56 2025-10-06T08:48:46 \r\n          30256 tokenizer_atari_alien_dev_lr_+         standard  16    COMPLETED   07:20:19 1-00:00:00 2025-10-05T16:54:26 2025-10-06T00:14:45 \r\n          30257 tokenizer_atari_amidar_dev_lr+         standard  16    COMPLETED   04:59:47 1-00:00:00 2025-10-05T17:52:56 2025-10-05T22:52:43 \r\n          30292 dynamics_atari_alien_dev_150k+         standard  16    COMPLETED   06:12:28 1-00:00:00 2025-10-06T01:50:39 2025-10-06T08:03:07 \r\n          30293 dynamics_atari_amidar_dev_150+         standard  16    COMPLETED   06:13:27 1-00:00:00 2025-10-06T01:50:39 2025-10-06T08:04:06 \r\n          30294 dynamics_atari_assault_dev_15+         standard  16    COMPLETED   06:14:02 1-00:00:00 2025-10-06T01:50:39 2025-10-06T08:04:41 \r\n          30295 dynamics_atari_asterix_dev_15+         standard  16    COMPLETED   06:14:34 1-00:00:00 2025-10-06T01:50:39 2025-10-06T08:05:13 \r\n          30296 dynamics_atari_bank_heist_dev+         standard  16    COMPLETED   06:16:00 1-00:00:00 2025-10-06T01:50:39 2025-10-06T08:06:39 \r\n          30297 dynamics_atari_battle_zone_de+         standard  16    COMPLETED   05:15:35 1-00:00:00 2025-10-06T03:22:29 2025-10-06T08:38:04 \r\n          30298 dynamics_atari_boxing_dev_150+         standard  16    COMPLETED   06:12:18 1-00:00:00 2025-10-06T05:53:25 2025-10-06T12:05:43 \r\n          30299 dynamics_atari_breakout_dev_1+         standard  16    COMPLETED   06:08:35 1-00:00:00 2025-10-06T05:53:25 2025-10-06T12:02:00 \r\n          30300 dynamics_atari_chopper_comman+         standard  16    COMPLETED   06:07:29 1-00:00:00 2025-10-06T05:53:25 2025-10-06T12:00:54 \r\n          30301 dynamics_atari_crazy_climber_+         standard  16    COMPLETED   06:09:03 1-00:00:00 2025-10-06T05:53:25 2025-10-06T12:02:28 \r\n          30302 dynamics_atari_demon_attack_d+         standard  16    COMPLETED   05:38:48 1-00:00:00 2025-10-06T05:59:25 2025-10-06T11:38:13 \r\n          30303 dynamics_atari_pong_dev_150k_+         standard  16    COMPLETED   05:38:18 1-00:00:00 2025-10-06T05:59:25 2025-10-06T11:37:43 \r\n          30280 tokenizer_atari_alien_dev_lr_+         standard  16    COMPLETED   04:22:42 1-00:00:00 2025-10-06T23:40:56 2025-10-07T04:03:38 \r\n          30281 tokenizer_atari_amidar_dev_lr+         standard  16    COMPLETED   19:11:52 1-00:00:00 2025-10-06T23:40:56 2025-10-07T18:52:48 \r\n          30284 tokenizer_atari_bank_heist_de+         standard  16    COMPLETED   05:08:15 1-00:00:00 2025-10-07T00:12:26 2025-10-07T05:20:41 \r\n          30285 tokenizer_atari_battle_zone_d+         standard  16    COMPLETED   19:19:43 1-00:00:00 2025-10-07T00:20:26 2025-10-07T19:40:09 \r\n          30286 tokenizer_atari_boxing_dev_lr+         standard  16    COMPLETED   12:04:22 1-00:00:00 2025-10-07T01:05:56 2025-10-07T13:10:18 \r\n          30288 tokenizer_atari_chopper_comma+         standard  16    COMPLETED   10:21:46 1-00:00:00 2025-10-07T01:42:57 2025-10-07T12:04:43 \r\n          30289 tokenizer_atari_crazy_climber+         standard  16    COMPLETED   22:11:51 1-00:00:00 2025-10-07T01:48:27 2025-10-08T00:00:18 \r\n          30291 tokenizer_atari_pong_dev_lr_3+         standard  16    COMPLETED   00:40:49 1-00:00:00 2025-10-07T01:48:27 2025-10-07T02:29:16 \r\n          30282 tokenizer_atari_assault_dev_l+         standard  16    COMPLETED   08:51:24 1-00:00:00 2025-10-07T23:38:34 2025-10-08T08:29:58 \r\n          30283 tokenizer_atari_asterix_dev_l+         standard  16    COMPLETED   03:38:40 1-00:00:00 2025-10-07T23:38:34 2025-10-08T03:17:14 \r\n          30287 tokenizer_atari_breakout_dev_+         standard  16    COMPLETED   04:43:13 1-00:00:00 2025-10-08T01:40:36 2025-10-08T06:23:49 \r\n          30496 dynamics_atari_alien_dev_fleu+         standard  16       FAILED   00:00:49 1-00:00:00 2025-10-08T09:34:03 2025-10-08T09:34:52 \r\n          30497 dynamics_atari_amidar_dev_fle+         standard  16       FAILED   00:01:11 1-00:00:00 2025-10-08T09:34:03 2025-10-08T09:35:14 \r\n          30498 dynamics_atari_assault_dev_fl+         standard  16       FAILED   00:00:46 1-00:00:00 2025-10-08T09:34:03 2025-10-08T09:34:49 \r\n          30499 dynamics_atari_asterix_dev_fl+         standard  16       FAILED   00:00:34 1-00:00:00 2025-10-08T09:34:03 2025-10-08T09:34:37 \r\n          30500 dynamics_atari_bank_heist_dev+         standard  16       FAILED   00:01:14 1-00:00:00 2025-10-08T09:34:03 2025-10-08T09:35:17 \r\n          30501 dynamics_atari_battle_zone_de+         standard  16       FAILED   00:00:51 1-00:00:00 2025-10-08T09:34:03 2025-10-08T09:34:54 \r\n          30502 dynamics_atari_boxing_dev_fle+         standard  16       FAILED   00:00:47 1-00:00:00 2025-10-08T09:34:03 2025-10-08T09:34:50 \r\n          30503 dynamics_atari_breakout_dev_f+         standard  16       FAILED   00:01:15 1-00:00:00 2025-10-08T09:34:03 2025-10-08T09:35:18 \r\n          30504 dynamics_atari_crazy_climber_+         standard  16       FAILED   00:00:51 1-00:00:00 2025-10-08T09:34:03 2025-10-08T09:34:54 \r\n          30505 dynamics_atari_alien_dev_fleu+         standard  16       FAILED   00:00:48 1-00:00:00 2025-10-08T09:50:33 2025-10-08T09:51:21 \r\n          30506 dynamics_atari_amidar_dev_fle+         standard  16       FAILED   00:00:47 1-00:00:00 2025-10-08T09:50:33 2025-10-08T09:51:20 \r\n          30507 dynamics_atari_assault_dev_fl+         standard  16       FAILED   00:00:32 1-00:00:00 2025-10-08T09:50:33 2025-10-08T09:51:05 \r\n          30508 dynamics_atari_asterix_dev_fl+         standard  16       FAILED   00:00:48 1-00:00:00 2025-10-08T09:50:33 2025-10-08T09:51:21 \r\n          30509 dynamics_atari_bank_heist_dev+         standard  16       FAILED   00:00:51 1-00:00:00 2025-10-08T09:50:33 2025-10-08T09:51:24 \r\n          30510 dynamics_atari_battle_zone_de+         standard  16       FAILED   00:00:34 1-00:00:00 2025-10-08T09:50:33 2025-10-08T09:51:07 \r\n          30511 dynamics_atari_boxing_dev_fle+         standard  16       FAILED   00:00:32 1-00:00:00 2025-10-08T09:50:33 2025-10-08T09:51:05 \r\n          30512 dynamics_atari_breakout_dev_f+         standard  16       FAILED   00:01:02 1-00:00:00 2025-10-08T09:50:33 2025-10-08T09:51:35 \r\n          30513 dynamics_atari_crazy_climber_+         standard  16       FAILED   00:00:31 1-00:00:00 2025-10-08T09:50:33 2025-10-08T09:51:04 \r\n          30514 dynamics_atari_alien_dev_fleu+         standard  16       FAILED   00:00:38 1-00:00:00 2025-10-08T09:55:33 2025-10-08T09:56:11 \r\n          30515 dynamics_atari_amidar_dev_fle+         standard  16       FAILED   00:00:46 1-00:00:00 2025-10-08T09:55:33 2025-10-08T09:56:19 \r\n          30516 dynamics_atari_assault_dev_fl+         standard  16       FAILED   00:00:48 1-00:00:00 2025-10-08T09:55:33 2025-10-08T09:56:21 \r\n          30517 dynamics_atari_asterix_dev_fl+         standard  16       FAILED   00:00:47 1-00:00:00 2025-10-08T09:55:33 2025-10-08T09:56:20 \r\n          30518 dynamics_atari_bank_heist_dev+         standard  16       FAILED   00:00:36 1-00:00:00 2025-10-08T09:55:33 2025-10-08T09:56:09 \r\n          30519 dynamics_atari_battle_zone_de+         standard  16       FAILED   00:00:38 1-00:00:00 2025-10-08T09:55:33 2025-10-08T09:56:11 \r\n          30520 dynamics_atari_boxing_dev_fle+         standard  16       FAILED   00:01:05 1-00:00:00 2025-10-08T09:55:33 2025-10-08T09:56:38 \r\n          30521 dynamics_atari_breakout_dev_f+         standard  16       FAILED   00:00:36 1-00:00:00 2025-10-08T09:55:33 2025-10-08T09:56:09 \r\n          30522 dynamics_atari_crazy_climber_+         standard  16       FAILED   00:00:46 1-00:00:00 2025-10-08T09:55:33 2025-10-08T09:56:19 \r\n          30523 dynamics_atari_alien_dev_fleu+         standard  16    COMPLETED   20:42:47 1-00:00:00 2025-10-08T10:09:33 2025-10-09T06:52:20 \r\n          30524 dynamics_atari_amidar_dev_fle+         standard  16    COMPLETED   20:42:23 1-00:00:00 2025-10-08T10:09:33 2025-10-09T06:51:56 \r\n          30525 dynamics_atari_assault_dev_fl+         standard  16    COMPLETED   20:47:41 1-00:00:00 2025-10-08T10:09:33 2025-10-09T06:57:14 \r\n          30526 dynamics_atari_asterix_dev_fl+         standard  16    COMPLETED   20:42:56 1-00:00:00 2025-10-08T10:09:33 2025-10-09T06:52:29 \r\n          30527 dynamics_atari_bank_heist_dev+         standard  16    COMPLETED   20:43:41 1-00:00:00 2025-10-08T10:09:33 2025-10-09T06:53:14 \r\n          30528 dynamics_atari_battle_zone_de+         standard  16    COMPLETED   20:46:16 1-00:00:00 2025-10-08T10:09:33 2025-10-09T06:55:49 \r\n          30529 dynamics_atari_boxing_dev_fle+         standard  16    COMPLETED   20:42:42 1-00:00:00 2025-10-08T10:09:33 2025-10-09T06:52:15 \r\n          30530 dynamics_atari_breakout_dev_f+         standard  16    COMPLETED   20:47:02 1-00:00:00 2025-10-08T10:09:33 2025-10-09T06:56:35 \r\n          30531 dynamics_atari_crazy_climber_+         standard  16    COMPLETED   20:44:45 1-00:00:00 2025-10-08T10:09:33 2025-10-09T06:54:18 \r\n          30290 tokenizer_atari_demon_attack_+         standard  16    COMPLETED   02:19:41 1-00:00:00 2025-10-09T01:43:42 2025-10-09T04:03:23 \r\n          31727 dynamics_atari_demon_attack_d+         standard  16    COMPLETED   20:47:08 1-00:00:00 2025-10-09T16:50:08 2025-10-10T13:37:16 \r\n          31761     sample_atari_alien_maskgit         standard  16    COMPLETED   00:01:13 1-00:00:00 2025-10-09T17:25:24 2025-10-09T17:26:37 \r\n          31762    sample_atari_amidar_maskgit         standard  16    COMPLETED   00:01:17 1-00:00:00 2025-10-09T17:25:24 2025-10-09T17:26:41 \r\n          31763   sample_atari_assault_maskgit         standard  16    COMPLETED   00:01:13 1-00:00:00 2025-10-09T17:25:24 2025-10-09T17:26:37 \r\n          31764   sample_atari_asterix_maskgit         standard  16    COMPLETED   00:01:12 1-00:00:00 2025-10-09T17:25:25 2025-10-09T17:26:37 \r\n          31765 sample_atari_bank_heist_maskg+         standard  16    COMPLETED   00:01:23 1-00:00:00 2025-10-09T17:25:25 2025-10-09T17:26:48 \r\n          31766 sample_atari_battle_zone_mask+         standard  16    COMPLETED   00:01:06 1-00:00:00 2025-10-09T17:25:25 2025-10-09T17:26:31 \r\n          31767  sample_atari_breakout_maskgit         standard  16    COMPLETED   00:01:01 1-00:00:00 2025-10-09T17:25:44 2025-10-09T17:26:45 \r\n          31768 sample_atari_crazy_climber_ma+         standard  16    COMPLETED   00:01:03 1-00:00:00 2025-10-09T17:25:44 2025-10-09T17:26:47 \r\n          31863 coinrun_sample_maskgit_mila_s+         standard  16       FAILED   00:00:14 1-00:00:00 2025-10-10T16:42:46 2025-10-10T16:43:00 \r\n          31864 coinrun_sample_maskgit_mila_s+         standard  16       FAILED   00:00:19 1-00:00:00 2025-10-10T16:42:46 2025-10-10T16:43:05 \r\n          31865 coinrun_sample_maskgit_mila_s+         standard  16       FAILED   00:00:13 1-00:00:00 2025-10-10T16:44:16 2025-10-10T16:44:29 \r\n          31866 coinrun_sample_maskgit_mila_s+         standard  16       FAILED   00:00:12 1-00:00:00 2025-10-10T16:44:46 2025-10-10T16:44:58 \r\n          31867 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:00 1-00:00:00 2025-10-10T16:47:46 2025-10-10T16:48:46 \r\n          31868 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:00:57 1-00:00:00 2025-10-10T16:52:16 2025-10-10T16:53:13 \r\n          31869 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:02 1-00:00:00 2025-10-10T17:01:46 2025-10-10T17:02:48 \r\n          31870 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:25 1-00:00:00 2025-10-10T17:02:16 2025-10-10T17:03:41 \r\n          31871 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:26 1-00:00:00 2025-10-10T17:08:16 2025-10-10T17:09:42 \r\n          31872 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:21 1-00:00:00 2025-10-10T17:28:46 2025-10-10T17:30:07 \r\n          31873 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:25 1-00:00:00 2025-10-10T17:29:46 2025-10-10T17:31:11 \r\n          31874 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:21 1-00:00:00 2025-10-10T17:34:16 2025-10-10T17:35:37 \r\n          31875 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:24 1-00:00:00 2025-10-10T17:36:46 2025-10-10T17:38:10 \r\n          31882 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:28 1-00:00:00 2025-10-10T18:14:16 2025-10-10T18:15:44 \r\n          31883 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:29 1-00:00:00 2025-10-10T18:16:46 2025-10-10T18:18:15 \r\n          32059 coinrun_sample_maskgit_mila_s+         standard  16       FAILED   00:00:01 1-00:00:00 2025-10-13T10:37:37 2025-10-13T10:37:38 \r\n          32060 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:11 1-00:00:00 2025-10-13T10:37:43 2025-10-13T10:38:54 \r\n          32061 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:00:57 1-00:00:00 2025-10-13T10:38:35 2025-10-13T10:39:32 \r\n          32062 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:03 1-00:00:00 2025-10-13T10:40:25 2025-10-13T10:41:28 \r\n          32063 coinrun_sample_maskgit_mila_s+         standard  16    COMPLETED   00:01:01 1-00:00:00 2025-10-13T10:40:55 2025-10-13T10:41:56 \r\n]0;franz.srambical@hai-login2:~/jafar",,terminal_output

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-bd85f04e-ab88-4199-88e6-b32c893cf6c31762531753584-2025_11_07-17.09.15.120/source.csv

@@ -0,0 +1,6 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,158,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"5:09:15 PM [info] Activating crowd-code\n5:09:15 PM [info] Recording started\n5:09:15 PM [info] Initializing git provider using file system watchers...\n5:09:15 PM [info] Git repository found\n5:09:15 PM [info] Git provider initialized successfully\n",Log,tab
+3,221,"extension-output-pdoom-org.crowd-code-#1-crowd-code",245,0,"5:09:15 PM [info] Initial git state: [object Object]\n",Log,content
+4,5051,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"Switched from branch 'connect-to-sglang' to 'main'",Log,git_branch_checkout
+5,5054,"extension-output-pdoom-org.crowd-code-#1-crowd-code",298,0,"5:09:20 PM [info] Branch checkout detected: connect-to-sglang -> main\n5:09:20 PM [info] Recording git checkout: Switched from branch 'connect-to-sglang' to 'main'\n5:09:20 PM [info] Resetting file cache due to branch checkout\n",Log,content
+6,11285,"TERMINAL",0,0,"",,terminal_focus

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-c66923f3-2a2a-4b19-880b-c1a8bfe1bf981753195775955-2025_07_22-16.49.43.384/source.csv

@@ -0,0 +1,79 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,".venv/lib/python3.10/site-packages/flax/linen/attention.py",0,0,"# Copyright 2024 The Flax Authors.\n#\n# Licensed under the Apache License, Version 2.0 (the ""License"");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n#     http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an ""AS IS"" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n\n""""""Attention core modules for Flax.""""""\nfrom __future__ import annotations\n\nimport functools\nimport inspect\nimport warnings\nfrom typing import Any, overload\nfrom collections.abc import Callable\n\nimport jax\nimport jax.numpy as jnp\nfrom jax import lax, random\n\nfrom flax.linen import initializers\nfrom flax.linen.dtypes import promote_dtype\nfrom flax.linen.linear import (\n  DenseGeneral,\n  default_kernel_init,\n)\nfrom flax.linen.module import Module, compact, merge_param\nfrom flax.linen.normalization import LayerNorm\nfrom flax.typing import (\n  Array,\n  PRNGKey,\n  Dtype,\n  Shape as Shape,\n  Initializer,\n  PrecisionLike,\n  DotGeneralT,\n)\n\n\ndef dot_product_attention_weights(\n    query: Array,\n    key: Array,\n    bias: Array | None = None,\n    mask: Array | None = None,\n    broadcast_dropout: bool = True,\n    dropout_rng: PRNGKey | None = None,\n    dropout_rate: float = 0.0,\n    deterministic: bool = False,\n    dtype: Dtype | None = None,\n    precision: PrecisionLike = None,\n    module: Module | None = None,\n    force_fp32_for_softmax: bool = False,\n    einsum_dot_general: Callable[..., Array] | None = None,\n    einsum: Callable[..., Array] | None = None,\n):\n  """"""Computes dot-product attention weights given query and key.\n\n  Used by :func:`dot_product_attention`, which is what you'll most likely use.\n  But if you want access to the attention weights for introspection, then\n  you can directly call this function and call einsum yourself.\n\n  Args:\n    query: queries for calculating attention with shape of ``[batch...,\n      q_length, num_heads, qk_depth_per_head]``.\n    key: keys for calculating attention with shape of ``[batch..., kv_length,\n      num_heads, qk_depth_per_head]``.\n    bias: bias for the attention weights. This should be broadcastable to the\n      shape ``[batch..., num_heads, q_length, kv_length]``. This can be used for\n      incorporating causal masks, padding masks, proximity bias, etc.\n    mask: mask for the attention weights. This should be broadcastable to the\n      shape ``[batch..., num_heads, q_length, kv_length]``. This can be used for\n      incorporating causal masks. Attention weights are masked out if their\n      corresponding mask value is ``False``.\n    broadcast_dropout: bool: use a broadcasted dropout along batch dims.\n    dropout_rng: JAX PRNGKey: to be used for dropout\n    dropout_rate: dropout rate\n    deterministic: bool, deterministic or not (to apply dropout)\n    dtype: the dtype of the computation (default: infer from inputs and params)\n    precision: numerical precision of the computation see ``jax.lax.Precision``\n      for details.\n    module: the Module that will sow the attention weights into the\n      'intermediates' collection. Remember to mark 'intermediates' as mutable\n      via ``mutable=['intermediates']`` in order to have that collection\n      returned. If ``module`` is None, the attention weights will not be sowed.\n    force_fp32_for_softmax: bool, whether to force the softmax to be computed in\n      fp32. This is useful for mixed-precision training where higher precision\n      is desired for numerical stability.\n    einsum_dot_general: the dot_general to use in einsum.\n    einsum: If unspecified, default `jnp.einsum` will be used. This argument is\n      mutually exclusive with `precision` and `einsum_dot_general`.\n\n  Raises:\n    ValueError: if both `precision`/`einsum_dot_general` and `einsum` are\n      specified.\n\n  Returns:\n    Output of shape ``[batch..., num_heads, q_length, kv_length]``.\n  """"""\n  if (precision or einsum_dot_general) and einsum:\n    raise ValueError(\n        'precision/einsum_dot_general and einsum are mutually exclusive. Please'\n        ' specify only one of them.'\n    )\n  if not einsum:\n    einsum = functools.partial(\n        jnp.einsum,\n        precision=precision,\n        _dot_general=einsum_dot_general\n        if einsum_dot_general\n        else jax.lax.dot_general,\n    )\n\n  query, key = promote_dtype(query, key, dtype=dtype)\n  dtype = query.dtype\n\n  assert query.ndim == key.ndim, 'q, k must have same rank.'\n  assert query.shape[:-3] == key.shape[:-3], 'q, k batch dims must match.'\n  assert query.shape[-2] == key.shape[-2], 'q, k num_heads must match.'\n  assert query.shape[-1] == key.shape[-1], 'q, k depths must match.'\n\n  # calculate attention matrix\n  depth = query.shape[-1]\n  query = query / jnp.sqrt(depth).astype(dtype)\n  # attn weight shape is (batch..., num_heads, q_length, kv_length)\n  attn_weights = einsum('...qhd,...khd->...hqk', query, key)\n\n  # apply attention bias: masking, dropout, proximity bias, etc.\n  if bias is not None:\n    attn_weights = attn_weights + bias\n  # apply attention mask\n  if mask is not None:\n    big_neg = jnp.finfo(dtype).min\n    attn_weights = jnp.where(mask, attn_weights, big_neg)\n\n  # normalize the attention weights\n  if force_fp32_for_softmax and dtype != jnp.float32:\n    attn_weights = jax.nn.softmax(attn_weights.astype(jnp.float32))\n  else:\n    attn_weights = jax.nn.softmax(attn_weights).astype(dtype)\n\n  if module:\n    module.sow('intermediates', 'attention_weights', attn_weights)\n\n  # apply attention dropout\n  if not deterministic and dropout_rate > 0.0:\n    keep_prob = 1.0 - dropout_rate\n    if broadcast_dropout:\n      # dropout is broadcast across the batch + head dimensions\n      dropout_shape = tuple([1] * (key.ndim - 2)) + attn_weights.shape[-2:]\n      keep = random.bernoulli(dropout_rng, keep_prob, dropout_shape)  # type: ignore\n    else:\n      keep = random.bernoulli(dropout_rng, keep_prob, attn_weights.shape)  # type: ignore\n    multiplier = keep.astype(dtype) / jnp.asarray(keep_prob, dtype=dtype)\n    attn_weights = attn_weights * multiplier\n\n  return attn_weights\n\n\ndef dot_product_attention(\n    query: Array,\n    key: Array,\n    value: Array,\n    bias: Array | None = None,\n    mask: Array | None = None,\n    broadcast_dropout: bool = True,\n    dropout_rng: PRNGKey | None = None,\n    dropout_rate: float = 0.0,\n    deterministic: bool = False,\n    dtype: Dtype | None = None,\n    precision: PrecisionLike = None,\n    module: Module | None = None,\n    force_fp32_for_softmax: bool = False,\n    einsum_dot_general: Callable[..., Array] | None = None,\n    qk_attn_weights_einsum: Callable[..., Array] | None = None,\n    attn_weights_value_einsum: Callable[..., Array] | None = None,\n):\n  """"""Computes dot-product attention given query, key, and value.\n\n  This is the core function for applying attention based on\n  https://arxiv.org/abs/1706.03762. It calculates the attention weights given\n  query and key and combines the values using the attention weights.\n\n  .. note::\n    ``query``, ``key``, ``value`` needn't have any batch dimensions.\n\n  Args:\n    query: queries for calculating attention with shape of ``[batch...,\n      q_length, num_heads, qk_depth_per_head]``.\n    key: keys for calculating attention with shape of ``[batch..., kv_length,\n      num_heads, qk_depth_per_head]``.\n    value: values to be used in attention with shape of ``[batch..., kv_length,\n      num_heads, v_depth_per_head]``.\n    bias: bias for the attention weights. This should be broadcastable to the\n      shape ``[batch..., num_heads, q_length, kv_length]``. This can be used for\n      incorporating causal masks, padding masks, proximity bias, etc.\n    mask: mask for the attention weights. This should be broadcastable to the\n      shape ``[batch..., num_heads, q_length, kv_length]``. This can be used for\n      incorporating causal masks. Attention weights are masked out if their\n      corresponding mask value is ``False``.\n    broadcast_dropout: bool: use a broadcasted dropout along batch dims.\n    dropout_rng: JAX PRNGKey: to be used for dropout\n    dropout_rate: dropout rate\n    deterministic: bool, deterministic or not (to apply dropout)\n    dtype: the dtype of the computation (default: infer from inputs)\n    precision: numerical precision of the computation see ``jax.lax.Precision`\n      for details.\n    module: the Module that will sow the attention weights into the\n      'intermediates' collection. Remember to mark 'intermediates' as mutable\n      via ``mutable=['intermediates']`` in order to have that collection\n      returned. If ``module`` is None, the attention weights will not be sowed.\n    force_fp32_for_softmax: bool, whether to force the softmax to be computed in\n      fp32. This is useful for mixed-precision training where higher precision\n      is desired for numerical stability.\n    einsum_dot_general: the dot_general to use in `jnp.einsum`.\n    qk_attn_weights_einsum: the einsum for computing the attention weights. When\n      unspecified, the default `jnp.einsum` will be used. This argument is\n      mutually exclusive with `precision` and `einsum_dot_general`.\n    attn_weights_value_einsum: the einsum for computing the product of the\n      attention weights and the values. When unspecified, the default\n      `jnp.einsum` will be used. This argument is mutually exclusive with\n      `precision` and `einsum_dot_general`.\n\n  Returns:\n    Output of shape ``[batch..., q_length, num_heads, v_depth_per_head]``.\n\n  Raises:\n    ValueError: if both `precision`/`einsum_dot_general` and\n    `qk_attn_weights_einsum`/`attn_weights_value_einsum` are\n      specified.\n  """"""\n  if (qk_attn_weights_einsum and not attn_weights_value_einsum) or (\n      not qk_attn_weights_einsum and attn_weights_value_einsum\n  ):\n    raise ValueError(\n        'qk_attn_weights_einsum and attn_weights_value_einsum must be specified'\n        ' together.'\n    )\n  if (precision or einsum_dot_general) and (\n      qk_attn_weights_einsum or attn_weights_value_einsum\n  ):\n    raise ValueError(\n        'precision/einsum_dot_general and'\n        ' qk_attn_weights_einsum/attn_weights_value_einsum are mutually'\n        ' exclusive. Please specify only one of them.'\n    )\n\n  query, key, value = promote_dtype(query, key, value, dtype=dtype)\n  dtype = query.dtype\n  assert key.ndim == query.ndim == value.ndim, 'q, k, v must have same rank.'\n  assert (\n    query.shape[:-3] == key.shape[:-3] == value.shape[:-3]\n  ), 'q, k, v batch dims must match.'\n  assert (\n    query.shape[-2] == key.shape[-2] == value.shape[-2]\n  ), 'q, k, v num_heads must match.'\n  assert key.shape[-3] == value.shape[-3], 'k, v lengths must match.'\n\n  # compute attention weights\n  attn_weights = dot_product_attention_weights(\n      query,\n      key,\n      bias,\n      mask,\n      broadcast_dropout,\n      dropout_rng,\n      dropout_rate,\n      deterministic,\n      dtype,\n      precision,\n      module,\n      force_fp32_for_softmax,\n      einsum_dot_general=einsum_dot_general,\n      einsum=qk_attn_weights_einsum,\n  )\n  if not attn_weights_value_einsum:\n    attn_weights_value_einsum = functools.partial(\n        jnp.einsum,\n        precision=precision,\n        _dot_general=einsum_dot_general\n        if einsum_dot_general\n        else jax.lax.dot_general,\n    )\n  # return weighted sum over values for each query position\n  return attn_weights_value_einsum(\n      '...hqk,...khd->...qhd',\n      attn_weights,\n      value,\n  )\n\n\nclass MultiHeadDotProductAttention(Module):\n  """"""Multi-head dot-product attention.\n\n  Example usage::\n\n    >>> import flax.linen as nn\n    >>> import jax\n\n    >>> layer = nn.MultiHeadDotProductAttention(num_heads=8, qkv_features=16)\n    >>> key1, key2, key3, key4, key5, key6 = jax.random.split(jax.random.key(0), 6)\n    >>> shape = (4, 3, 2, 5)\n    >>> q, k, v = jax.random.uniform(key1, shape), jax.random.uniform(key2, shape), jax.random.uniform(key3, shape)\n    >>> variables = layer.init(jax.random.key(0), q)\n\n    >>> # different inputs for inputs_q, inputs_k and inputs_v\n    >>> out = layer.apply(variables, q, k, v)\n    >>> # equivalent to layer.apply(variables, inputs_q=q, inputs_k=k, inputs_v=k)\n    >>> out = layer.apply(variables, q, k)\n    >>> # equivalent to layer.apply(variables, inputs_q=q, inputs_k=q) and layer.apply(variables, inputs_q=q, inputs_k=q, inputs_v=q)\n    >>> out = layer.apply(variables, q)\n\n    >>> attention_kwargs = dict(\n    ...     num_heads=8,\n    ...     qkv_features=16,\n    ...     kernel_init=nn.initializers.ones,\n    ...     bias_init=nn.initializers.zeros,\n    ...     dropout_rate=0.5,\n    ...     deterministic=False,\n    ...     )\n    >>> class Module(nn.Module):\n    ...   attention_kwargs: dict\n    ...\n    ...   @nn.compact\n    ...   def __call__(self, x, dropout_rng=None):\n    ...     out1 = nn.MultiHeadDotProductAttention(**self.attention_kwargs)(x, dropout_rng=dropout_rng)\n    ...     out2 = nn.MultiHeadDotProductAttention(**self.attention_kwargs)(x, dropout_rng=dropout_rng)\n    ...     return out1, out2\n    >>> module = Module(attention_kwargs)\n    >>> variables = module.init({'params': key1, 'dropout': key2}, q)\n\n    >>> # out1 and out2 are different.\n    >>> out1, out2 = module.apply(variables, q, rngs={'dropout': key3})\n    >>> # out3 and out4 are different.\n    >>> # out1 and out3 are different. out2 and out4 are different.\n    >>> out3, out4 = module.apply(variables, q, rngs={'dropout': key4})\n    >>> # out1 and out2 are the same.\n    >>> out1, out2 = module.apply(variables, q, dropout_rng=key5)\n    >>> # out1 and out2 are the same as out3 and out4.\n    >>> # providing a `dropout_rng` arg will take precedence over the `rngs` arg in `.apply`\n    >>> out3, out4 = module.apply(variables, q, rngs={'dropout': key6}, dropout_rng=key5)\n\n  Attributes:\n    num_heads: Number of attention heads. Features (i.e. inputs_q.shape[-1])\n      should be divisible by the number of heads.\n    dtype: The dtype of the computation (default: infer from inputs and params)\n    param_dtype: The dtype passed to parameter initializers (default: float32)\n    qkv_features: Dimension of the key, query, and value.\n    out_features: Dimension of the last projection\n    broadcast_dropout: Use a broadcasted dropout along batch dims.\n    dropout_rate: Dropout rate.\n    deterministic: If False, the attention weight is masked randomly using\n      dropout, whereas if True, the attention weights are deterministic.\n    precision: Numerical precision of the computation see ``jax.lax.Precision``\n      for details.\n    kernel_init: Initializer for the kernel of the Dense layers.\n    out_kernel_init: Optional Initializer for the kernel of the output Dense layer,\n      if None, ``kernel_init`` will be used.\n    bias_init: Initializer for the bias of the Dense layers.\n    out_bias_init: Optional Initializer for the bias of the output Dense layer,\n      if None, ``bias_init`` will be used.\n    use_bias: Whether pointwise QKVO dense transforms use bias.\n    attention_fn: dot_product_attention or compatible function. Accepts query,\n      key, value, and returns output of shape ``[bs, dim1, dim2, ..., dimN,,\n      num_heads, value_channels]``\n    decode: Whether to prepare and use an autoregressive cache.\n    normalize_qk: Should QK normalization be applied (arxiv.org/abs/2302.05442).\n    qk_attn_weights_einsum_cls: factory function to create the einsum for\n      computing the attention weights.\n    attn_weights_value_einsum_cls: factory function to create the einsum for\n      computing the product of the attention weights and the values.\n  """"""\n\n  num_heads: int\n  dtype: Dtype | None = None\n  param_dtype: Dtype = jnp.float32\n  qkv_features: int | None = None\n  out_features: int | None = None\n  broadcast_dropout: bool = True\n  dropout_rate: float = 0.0\n  deterministic: bool | None = None\n  precision: PrecisionLike = None\n  kernel_init: Initializer = default_kernel_init\n  out_kernel_init: Initializer | None = None\n  bias_init: Initializer = initializers.zeros_init()\n  out_bias_init: Initializer | None = None\n  use_bias: bool = True\n  attention_fn: Callable[..., Array] = dot_product_attention\n  decode: bool = False\n  normalize_qk: bool = False\n  force_fp32_for_softmax: bool = False\n  # Deprecated, will be removed.\n  qkv_dot_general: DotGeneralT | None = None\n  out_dot_general: DotGeneralT | None = None\n  qkv_dot_general_cls: Any = None\n  out_dot_general_cls: Any = None\n  qk_attn_weights_einsum_cls: Callable[..., Callable[..., Array]] | None = None\n  attn_weights_value_einsum_cls: Callable[..., Callable[..., Array]] | None = (\n      None\n  )\n\n  @overload\n  def __call__(\n    self,\n    inputs_q: Array,\n    inputs_k: Array | None = None,\n    inputs_v: Array | None = None,\n    *,\n    mask: Array | None = None,\n    deterministic: bool | None = None,\n    dropout_rng: PRNGKey | None = None,\n    sow_weights: bool = False,\n  ):\n    ...\n\n  @overload\n  def __call__(\n    self,\n    inputs_q: Array,\n    *,\n    inputs_kv: Array | None = None,\n    mask: Array | None = None,\n    deterministic: bool | None = None,\n    dropout_rng: PRNGKey | None = None,\n    sow_weights: bool = False,\n  ):\n    ...\n\n  @compact\n  def __call__(\n    self,\n    inputs_q: Array,\n    inputs_k: Array | None = None,\n    inputs_v: Array | None = None,\n    *,\n    inputs_kv: Array | None = None,\n    mask: Array | None = None,\n    deterministic: bool | None = None,\n    dropout_rng: PRNGKey | None = None,\n    sow_weights: bool = False,\n  ):\n    """"""Applies multi-head dot product attention on the input data.\n\n    Projects the inputs into multi-headed query, key, and value vectors,\n    applies dot-product attention and project the results to an output vector.\n\n    If both inputs_k and inputs_v are None, they will both copy the value of\n    inputs_q (self attention).\n    If only inputs_v is None, it will copy the value of inputs_k.\n\n    Args:\n      inputs_q: input queries of shape ``[batch_sizes..., length, features]``.\n      inputs_k: key of shape ``[batch_sizes..., length, features]``. If None,\n        inputs_k will copy the value of inputs_q.\n      inputs_v: values of shape ``[batch_sizes..., length, features]``. If None,\n        inputs_v will copy the value of inputs_k.\n      inputs_kv: key/values of shape ``[batch_sizes..., length, features]``. If\n        None, inputs_kv will copy the value of inputs_q. This arg will be\n        deprecated soon. Use inputs_k and inputs_v instead.\n      mask: attention mask of shape ``[batch_sizes..., num_heads, query_length,\n        key/value_length]``. Attention weights are masked out if their\n        corresponding mask value is ``False``.\n      deterministic: if false, the attention weight is masked randomly using\n        dropout, whereas if true, the attention weights are deterministic.\n      dropout_rng: optional rng key to pass to the attention layer's dropout\n        mask. Otherwise, self.make_rng('dropout') is used instead.\n      sow_weights: if ``True``, the attention weights are sowed into the\n        'intermediates' collection. Remember to mark 'intermediates' as\n        mutable via ``mutable=['intermediates']`` in order to have that\n        collection returned.\n\n    Returns:\n      output of shape ``[batch_sizes..., length, features]``.\n    """"""\n    if inputs_kv is not None:\n      if inputs_k is not None or inputs_v is not None:\n        raise ValueError(\n          'If either `inputs_k` or `inputs_v` is not None, '\n          '`inputs_kv` must be None. If `inputs_kv` is not None, both `inputs_k` '\n          'and `inputs_v` must be None. We recommend using `inputs_k` and '\n          '`inputs_v` args, since `inputs_kv` will be deprecated soon. See '\n          'https://github.com/google/flax/discussions/3389 for more '\n          'information.'\n        )\n      inputs_k = inputs_v = inputs_kv\n      warnings.warn(\n        'The inputs_kv arg will be deprecated soon. '\n        'Use inputs_k and inputs_v instead. See '\n        'https://github.com/google/flax/discussions/3389 '\n        'for more information.',\n        DeprecationWarning,\n      )\n    else:\n      if inputs_k is None:\n        if inputs_v is not None:\n          raise ValueError(\n            '`inputs_k` cannot be None if `inputs_v` is not None. '\n            'To have both `inputs_k` and `inputs_v` be the same value, pass in the '\n            'value to `inputs_k` and leave `inputs_v` as None.'\n          )\n        inputs_k = inputs_q\n      if inputs_v is None:\n        inputs_v = inputs_k\n      elif inputs_v.shape[-1] == inputs_v.shape[-2]:\n        warnings.warn(\n          f'You are passing an array of shape {inputs_v.shape} '\n          'to the `inputs_v` arg, when you may have intended '\n          'to pass it to the `mask` arg. As of Flax version '\n          '0.7.4, the function signature of '\n          ""MultiHeadDotProductAttention's `__call__` method ""\n          'has changed to `__call__(inputs_q, inputs_k=None, '\n          'inputs_v=None, *, inputs_kv=None, mask=None, '\n          'deterministic=None)`. Use the kwarg `mask` instead. '\n          'See https://github.com/google/flax/discussions/3389 '\n          'and read the docstring for more information.',\n          DeprecationWarning,\n        )\n\n    features = self.out_features or inputs_q.shape[-1]\n    qkv_features = self.qkv_features or inputs_q.shape[-1]\n    assert qkv_features % self.num_heads == 0, (\n      f'Memory dimension ({qkv_features}) must be divisible by number of'\n      f' heads ({self.num_heads}).'\n    )\n    head_dim = qkv_features // self.num_heads\n\n    dense = functools.partial(\n      DenseGeneral,\n      axis=-1,\n      dtype=self.dtype,\n      param_dtype=self.param_dtype,\n      features=(self.num_heads, head_dim),\n      kernel_init=self.kernel_init,\n      bias_init=self.bias_init,\n      use_bias=self.use_bias,\n      precision=self.precision,\n      dot_general=self.qkv_dot_general,\n      dot_general_cls=self.qkv_dot_general_cls,\n    )\n    # project inputs_q to multi-headed q/k/v\n    # dimensions are then [batch..., length, n_heads, n_features_per_head]\n    query, key, value = (\n      dense(name='query')(inputs_q),\n      dense(name='key')(inputs_k),\n      dense(name='value')(inputs_v),\n    )\n\n    if self.normalize_qk:\n      # Normalizing query and key projections stabilizes training with higher\n      # LR. See ViT-22B paper http://arxiv.org/abs/2302.05442 for analysis.\n      query = LayerNorm(\n        name='query_ln',\n        use_bias=False,\n        dtype=self.dtype,\n        param_dtype=self.param_dtype,\n      )(query)  # type: ignore[call-arg]\n      key = LayerNorm(\n        name='key_ln',\n        use_bias=False,\n        dtype=self.dtype,\n        param_dtype=self.param_dtype,\n      )(key)  # type: ignore[call-arg]\n\n    # During fast autoregressive decoding, we feed one position at a time,\n    # and cache the keys and values step by step.\n    if self.decode:\n      # detect if we're initializing by absence of existing cache data.\n      is_initialized = self.has_variable('cache', 'cached_key')\n      cached_key = self.variable(\n        'cache', 'cached_key', jnp.zeros, key.shape, key.dtype\n      )\n      cached_value = self.variable(\n        'cache', 'cached_value', jnp.zeros, value.shape, value.dtype\n      )\n      cache_index = self.variable(\n        'cache', 'cache_index', lambda: jnp.array(0, dtype=jnp.int32)\n      )\n      if is_initialized:\n        (\n          *batch_dims,\n          max_length,\n          num_heads,\n          depth_per_head,\n        ) = cached_key.value.shape\n        # shape check of cached keys against query input\n        expected_shape = tuple(batch_dims) + (1, num_heads, depth_per_head)\n        if expected_shape != query.shape:\n          raise ValueError(\n            'Autoregressive cache shape error, '\n            'expected query shape %s instead got %s.'\n            % (expected_shape, query.shape)\n          )\n        # update key, value caches with our new 1d spatial slices\n        cur_index = cache_index.value\n        zero = jnp.array(0, dtype=lax.dtype(cur_index.dtype))\n        indices: tuple[int | jax.Array, ...] = (zero,) * len(\n          batch_dims\n        ) + (\n          cur_index,\n          zero,\n          zero,\n        )\n        key = lax.dynamic_update_slice(cached_key.value, key, indices)\n        value = lax.dynamic_update_slice(cached_value.value, value, indices)\n        cached_key.value = key\n        cached_value.value = value\n        cache_index.value = cache_index.value + 1\n        # causal mask for cached decoder self-attention:\n        # our single query position should only attend to those key\n        # positions that have already been generated and cached,\n        # not the remaining zero elements.\n        mask = combine_masks(\n          mask,\n          jnp.broadcast_to(\n            jnp.arange(max_length) <= cur_index,\n            tuple(batch_dims) + (1, 1, max_length),\n          ),\n        )\n\n    if (\n      self.dropout_rate > 0.0\n    ):  # Require `deterministic` only if using dropout.\n      m_deterministic = merge_param(\n        'deterministic', self.deterministic, deterministic\n      )\n      if not m_deterministic and dropout_rng is None:\n        dropout_rng = self.make_rng('dropout')\n    else:\n      m_deterministic = True\n\n    # `qk_attn_weights_einsum` and `attn_weights_value_einsum` are optional\n    # arguments that can be used to override the default `jnp.einsum`. They\n    # exist for quantized einsum support in AQT.\n    qk_attn_weights_einsum = (\n        self.qk_attn_weights_einsum_cls()\n        if self.qk_attn_weights_einsum_cls\n        else None\n    )\n    attn_weights_value_einsum = (\n        self.attn_weights_value_einsum_cls()\n        if self.attn_weights_value_einsum_cls\n        else None\n    )\n    # apply attention\n    attn_args = (query, key, value)\n    # This kwargs list match the default nn.dot_product_attention.\n    # For custom `attention_fn`s, invalid kwargs will be filtered.\n    attn_kwargs = dict(\n      mask=mask,\n      dropout_rng=dropout_rng,\n      dropout_rate=self.dropout_rate,\n      broadcast_dropout=self.broadcast_dropout,\n      deterministic=m_deterministic,\n      dtype=self.dtype,\n      precision=self.precision,\n      force_fp32_for_softmax=self.force_fp32_for_softmax,\n      qk_attn_weights_einsum=qk_attn_weights_einsum,\n      attn_weights_value_einsum=attn_weights_value_einsum,\n    )\n    attn_kwargs = {\n        k: v\n        for k, v in attn_kwargs.items()\n        if k in inspect.signature(self.attention_fn).parameters\n    }\n    if sow_weights:\n      x = self.attention_fn(*attn_args, **attn_kwargs, module=self)\n    else:\n      x = self.attention_fn(*attn_args, **attn_kwargs)\n    # back to the original inputs dimensions\n    out = DenseGeneral(\n      features=features,\n      axis=(-2, -1),\n      kernel_init=self.out_kernel_init or self.kernel_init,\n      bias_init=self.out_bias_init or self.bias_init,\n      use_bias=self.use_bias,\n      dtype=self.dtype,\n      param_dtype=self.param_dtype,\n      precision=self.precision,\n      dot_general=self.out_dot_general,\n      dot_general_cls=self.out_dot_general_cls,\n      name='out',  # type: ignore[call-arg]\n    )(x)\n    return out\n\n\nclass MultiHeadAttention(MultiHeadDotProductAttention):\n  """"""Multi-head dot-product attention.\n  Alias for ``MultiHeadDotProductAttention``.\n\n  **NOTE**: ``MultiHeadAttention`` is a wrapper of ``MultiHeadDotProductAttention``,\n  and so their implementations are identical. However ``MultiHeadAttention`` layers\n  will, by default, be named ``MultiHeadAttention_{index}``, whereas ``MultiHeadDotProductAttention``\n  will be named ``MultiHeadDotProductAttention_{index}``. Therefore, this could affect\n  checkpointing, param collection names and RNG threading (since the layer name is\n  used when generating new RNG's) within the module.\n\n  Example usage::\n\n    >>> import flax.linen as nn\n    >>> import jax\n\n    >>> layer = nn.MultiHeadAttention(num_heads=8, qkv_features=16)\n    >>> key1, key2, key3, key4, key5, key6 = jax.random.split(jax.random.key(0), 6)\n    >>> shape = (4, 3, 2, 5)\n    >>> q, k, v = jax.random.uniform(key1, shape), jax.random.uniform(key2, shape), jax.random.uniform(key3, shape)\n    >>> variables = layer.init(jax.random.key(0), q)\n\n    >>> # different inputs for inputs_q, inputs_k and inputs_v\n    >>> out = layer.apply(variables, q, k, v)\n    >>> # equivalent to layer.apply(variables, inputs_q=q, inputs_k=k, inputs_v=k)\n    >>> out = layer.apply(variables, q, k)\n    >>> # equivalent to layer.apply(variables, inputs_q=q, inputs_k=q) and layer.apply(variables, inputs_q=q, inputs_k=q, inputs_v=q)\n    >>> out = layer.apply(variables, q)\n\n    >>> attention_kwargs = dict(\n    ...     num_heads=8,\n    ...     qkv_features=16,\n    ...     kernel_init=nn.initializers.ones,\n    ...     bias_init=nn.initializers.zeros,\n    ...     dropout_rate=0.5,\n    ...     deterministic=False,\n    ...     )\n    >>> class Module(nn.Module):\n    ...   attention_kwargs: dict\n    ...\n    ...   @nn.compact\n    ...   def __call__(self, x, dropout_rng=None):\n    ...     out1 = nn.MultiHeadAttention(**self.attention_kwargs)(x, dropout_rng=dropout_rng)\n    ...     out2 = nn.MultiHeadAttention(**self.attention_kwargs)(x, dropout_rng=dropout_rng)\n    ...     return out1, out2\n    >>> module = Module(attention_kwargs)\n    >>> variables = module.init({'params': key1, 'dropout': key2}, q)\n\n    >>> # out1 and out2 are different.\n    >>> out1, out2 = module.apply(variables, q, rngs={'dropout': key3})\n    >>> # out3 and out4 are different.\n    >>> # out1 and out3 are different. out2 and out4 are different.\n    >>> out3, out4 = module.apply(variables, q, rngs={'dropout': key4})\n    >>> # out1 and out2 are the same.\n    >>> out1, out2 = module.apply(variables, q, dropout_rng=key5)\n    >>> # out1 and out2 are the same as out3 and out4.\n    >>> # providing a `dropout_rng` arg will take precedence over the `rngs` arg in `.apply`\n    >>> out3, out4 = module.apply(variables, q, rngs={'dropout': key6}, dropout_rng=key5)\n\n  Attributes:\n    num_heads: number of attention heads. Features (i.e. inputs_q.shape[-1])\n      should be divisible by the number of heads.\n    dtype: the dtype of the computation (default: infer from inputs and params)\n    param_dtype: the dtype passed to parameter initializers (default: float32)\n    qkv_features: dimension of the key, query, and value.\n    out_features: dimension of the last projection\n    broadcast_dropout: bool: use a broadcasted dropout along batch dims.\n    dropout_rate: dropout rate\n    deterministic: if false, the attention weight is masked randomly using\n      dropout, whereas if true, the attention weights are deterministic.\n    precision: numerical precision of the computation see ``jax.lax.Precision``\n      for details.\n    kernel_init: initializer for the kernel of the Dense layers.\n    bias_init: initializer for the bias of the Dense layers.\n    use_bias: bool: whether pointwise QKVO dense transforms use bias.\n    attention_fn: dot_product_attention or compatible function. Accepts query,\n      key, value, and returns output of shape ``[bs, dim1, dim2, ..., dimN,,\n      num_heads, value_channels]``\n    decode: whether to prepare and use an autoregressive cache.\n    normalize_qk: should QK normalization be applied (arxiv.org/abs/2302.05442).\n  """"""\n\n\nclass SelfAttention(MultiHeadDotProductAttention):\n  """"""Self-attention special case of multi-head dot-product attention.\n  This layer is deprecated in favor of ``MultiHeadDotProductAttention``.\n\n  Example usage::\n    >>> import flax.linen as nn\n    >>> import jax, jax.numpy as jnp\n    >>> layer = nn.MultiHeadDotProductAttention(num_heads=8, qkv_features=16)\n    >>> variables = layer.init(jax.random.key(0), jnp.ones((4, 3, 2, 5)))\n  """"""\n\n  @compact\n  def __call__(  # type: ignore\n    self,\n    inputs_q: Array,\n    mask: Array | None = None,\n    deterministic: bool | None = None,\n    dropout_rng: PRNGKey | None = None,\n    sow_weights: bool = False,\n  ):\n    """"""Applies multi-head dot product self-attention on the input data.\n\n    Projects the inputs into multi-headed query, key, and value vectors,\n    applies dot-product attention and project the results to an output vector.\n\n    Args:\n      inputs_q: input queries of shape ``[batch_sizes..., length, features]``.\n      mask: attention mask of shape ``[batch_sizes..., num_heads, query_length,\n        key/value_length]``. Attention weights are masked out if their\n        corresponding mask value is ``False``.\n      deterministic: if false, the attention weight is masked randomly using\n        dropout, whereas if true, the attention weights are deterministic.\n\n    Returns:\n      output of shape ``[batch_sizes..., length, features]``.\n    """"""\n    warnings.warn(\n      'SelfAttention will be deprecated soon. Use '\n      '`MultiHeadDotProductAttention.__call__(inputs_q)` instead. '\n      'See https://github.com/google/flax/discussions/3389 '\n      'for more information.',\n      DeprecationWarning,\n    )\n    return super().__call__(\n      inputs_q,\n      mask=mask,\n      deterministic=deterministic,\n      dropout_rng=dropout_rng,\n      sow_weights=sow_weights,\n    )\n\n\n# mask-making utility functions\n\n\ndef make_attention_mask(\n  query_input: Array,\n  key_input: Array,\n  pairwise_fn: Callable[..., Any] = jnp.multiply,\n  extra_batch_dims: int = 0,\n  dtype: Dtype = jnp.float32,\n):\n  """"""Mask-making helper for attention weights.\n\n  In case of 1d inputs (i.e., ``[batch..., len_q]``, ``[batch..., len_kv]``, the\n  attention weights will be ``[batch..., heads, len_q, len_kv]`` and this\n  function will produce ``[batch..., 1, len_q, len_kv]``.\n\n  Args:\n    query_input: a batched, flat input of query_length size\n    key_input: a batched, flat input of key_length size\n    pairwise_fn: broadcasting elementwise comparison function\n    extra_batch_dims: number of extra batch dims to add singleton axes for, none\n      by default\n    dtype: mask return dtype\n\n  Returns:\n    A ``[batch..., 1, len_q, len_kv]`` shaped mask for 1d attention.\n  """"""\n  mask = pairwise_fn(\n    jnp.expand_dims(query_input, axis=-1), jnp.expand_dims(key_input, axis=-2)\n  )\n  mask = jnp.expand_dims(mask, axis=-3)\n  mask = jnp.expand_dims(mask, axis=tuple(range(extra_batch_dims)))\n  return mask.astype(dtype)\n\n\ndef make_causal_mask(\n  x: Array, extra_batch_dims: int = 0, dtype: Dtype = jnp.float32\n) -> Array:\n  """"""Make a causal mask for self-attention.\n\n  In case of 1d inputs (i.e., ``[batch..., len]``, the self-attention weights\n  will be ``[batch..., heads, len, len]`` and this function will produce a\n  causal mask of shape ``[batch..., 1, len, len]``.\n\n  Args:\n    x: input array of shape ``[batch..., len]``\n    extra_batch_dims: number of batch dims to add singleton axes for, none by\n      default\n    dtype: mask return dtype\n\n  Returns:\n    A ``[batch..., 1, len, len]`` shaped causal mask for 1d attention.\n  """"""\n  idxs = jnp.broadcast_to(jnp.arange(x.shape[-1], dtype=jnp.int32), x.shape)\n  return make_attention_mask(\n    idxs,\n    idxs,\n    jnp.greater_equal,\n    extra_batch_dims=extra_batch_dims,\n    dtype=dtype,\n  )\n\n\ndef combine_masks(\n  *masks: Array | None, dtype: Dtype = jnp.float32\n) -> Array | None:\n  """"""Combine attention masks.\n\n  Args:\n    *masks: set of attention mask arguments to combine, some can be None.\n    dtype: dtype for the returned mask.\n\n  Returns:\n    Combined mask, reduced by logical and, returns None if no masks given.\n  """"""\n  masks_list = [m for m in masks if m is not None]\n  if not masks_list:\n    return None\n  assert all(\n    map(lambda x: x.ndim == masks_list[0].ndim, masks_list)\n  ), f'masks must have same rank: {tuple(map(lambda x: x.ndim, masks_list))}'\n  mask, *other_masks = masks_list\n  for other_mask in other_masks:\n    mask = jnp.logical_and(mask, other_mask)\n  return mask.astype(dtype)\n",python,tab
+2,162,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"4:49:43 PM [info] Activating crowd-code\n4:49:43 PM [info] Recording started\n4:49:43 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,204,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"4:49:43 PM [info] Git repository found\n4:49:43 PM [info] Git provider initialized successfully\n4:49:43 PM [info] Initial git state: [object Object]\n",Log,content
+4,972,".venv/lib/python3.10/site-packages/flax/linen/attention.py",0,0,"",python,tab
+5,40295,".venv/lib/python3.10/site-packages/flax/linen/attention.py",13921,0,"",python,selection_command
+6,40716,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16444,0,"",python,selection_command
+7,42300,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16399,0,"",python,selection_command
+8,42558,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16366,0,"",python,selection_command
+9,42580,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16327,0,"",python,selection_command
+10,42613,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16298,0,"",python,selection_command
+11,42645,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16275,0,"",python,selection_command
+12,42673,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16214,0,"",python,selection_command
+13,42711,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16190,0,"",python,selection_command
+14,42743,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16147,0,"",python,selection_command
+15,42777,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16094,0,"",python,selection_command
+16,42810,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16049,0,"",python,selection_command
+17,42844,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16000,0,"",python,selection_command
+18,42878,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15966,0,"",python,selection_command
+19,42911,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15930,0,"",python,selection_command
+20,42945,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15902,0,"",python,selection_command
+21,42977,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15869,0,"",python,selection_command
+22,43011,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15835,0,"",python,selection_command
+23,43044,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15801,0,"",python,selection_command
+24,43077,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15766,0,"",python,selection_command
+25,43110,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15737,0,"",python,selection_command
+26,43143,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15720,0,"",python,selection_command
+27,43177,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15717,0,"",python,selection_command
+28,43229,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15713,0,"",python,selection_command
+29,43246,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15644,0,"",python,selection_command
+30,43278,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15567,0,"",python,selection_command
+31,43311,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15528,0,"",python,selection_command
+32,43344,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15454,0,"",python,selection_command
+33,43378,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15528,0,"",python,selection_command
+34,43635,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15567,0,"",python,selection_command
+35,43664,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15644,0,"",python,selection_command
+36,43694,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15713,0,"",python,selection_command
+37,43727,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15717,0,"",python,selection_command
+38,43761,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15720,0,"",python,selection_command
+39,43793,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15737,0,"",python,selection_command
+40,43827,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15766,0,"",python,selection_command
+41,43861,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15801,0,"",python,selection_command
+42,43895,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15835,0,"",python,selection_command
+43,43930,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15869,0,"",python,selection_command
+44,43962,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15902,0,"",python,selection_command
+45,43996,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15930,0,"",python,selection_command
+46,44030,".venv/lib/python3.10/site-packages/flax/linen/attention.py",15966,0,"",python,selection_command
+47,44066,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16000,0,"",python,selection_command
+48,44782,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16049,0,"",python,selection_command
+49,45035,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16094,0,"",python,selection_command
+50,45063,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16147,0,"",python,selection_command
+51,45095,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16190,0,"",python,selection_command
+52,45210,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16214,0,"",python,selection_command
+53,45384,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16275,0,"",python,selection_command
+54,45643,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16298,0,"",python,selection_command
+55,45666,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16327,0,"",python,selection_command
+56,45695,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16366,0,"",python,selection_command
+57,45732,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16399,0,"",python,selection_command
+58,45761,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16444,0,"",python,selection_command
+59,45795,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16489,0,"",python,selection_command
+60,45934,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16523,0,"",python,selection_command
+61,46084,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16557,0,"",python,selection_command
+62,50566,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16523,0,"",python,selection_command
+63,50816,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16489,0,"",python,selection_command
+64,50854,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16444,0,"",python,selection_command
+65,50871,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16399,0,"",python,selection_command
+66,50900,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16366,0,"",python,selection_command
+67,50936,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16327,0,"",python,selection_command
+68,51045,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16298,0,"",python,selection_command
+69,51198,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16275,0,"",python,selection_command
+70,51334,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16214,0,"",python,selection_command
+71,51454,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16190,0,"",python,selection_command
+72,51622,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16147,0,"",python,selection_command
+73,51714,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16190,0,"",python,selection_command
+74,51898,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16214,0,"",python,selection_command
+75,52063,".venv/lib/python3.10/site-packages/flax/linen/attention.py",16275,0,"",python,selection_command
+76,53668,".venv/lib/python3.10/site-packages/flax/linen/attention.py",22967,0,"",python,selection_command
+77,55696,".venv/lib/python3.10/site-packages/flax/linen/attention.py",22987,0,"",python,selection_command
+78,55812,".venv/lib/python3.10/site-packages/flax/linen/attention.py",23059,0,"",python,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-e5cda317-734f-4677-bc25-61ed75c3a5fb1767946586547-2026_01_09-09.16.35.258/source.csv

@@ -0,0 +1,220 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"src/recording.ts",0,0,"/**\n * Recording Orchestrator for crowd-code 2.0\n * Integrates viewport, terminal, filesystem, and deduplication modules\n * Implements the observation-action paradigm\n */\n\nimport * as fs from 'node:fs'\nimport * as path from 'node:path'\nimport * as vscode from 'vscode'\nimport axios from 'axios'\nimport { createTwoFilesPatch } from 'diff'\nimport { hasConsent } from './consent'\nimport {\n    notificationWithProgress,\n    generateBaseFilePath,\n    formatDisplayTime,\n    getExportPath,\n    logToOutput,\n    getConfig,\n    addToGitignore,\n} from './utilities'\nimport type {\n\tRecordingState,\n\tRecordingEvent,\n\tRecordingSession,\n\tObservation,\n\tAction,\n\tEditAction,\n\tSelectionAction,\n\tTabSwitchAction,\n\tTerminalFocusAction,\n\tTerminalCommandAction,\n\tTerminalOutputAction,\n\tFileChangeAction,\n\tScrollAction,\n} from './types'\nimport { extContext, statusBarItem, actionsProvider } from './extension'\nimport {\n\tcaptureObservation,\n\tresetObservationState,\n\tresetTerminalState,\n\tinitializeViewportCapture,\n\tinitializeTerminalCapture,\n\tinitializeFilesystemWatcher,\n\tresetFilesystemState,\n} from './capture'\nimport { getRecentGitOperation } from './gitProvider'\n\nexport const recording: RecordingState = {\n\tisRecording: false,\n\tstartDateTime: null,\n\tendDateTime: null,\n\tsequence: 0,\n\tsessionId: vscode.env.sessionId,\n\tevents: [],\n}\n\n\nexport const commands = {\n    openSettings: 'crowd-code.openSettings',\n    startRecording: 'crowd-code.startRecording',\n    stopRecording: 'crowd-code.stopRecording',\n    panicButton: 'crowd-code.panicButton',\n}\n\n\nlet intervalId: NodeJS.Timeout | null = null\nlet uploadIntervalId: NodeJS.Timeout | null = null\nlet timer = 0\nlet previousFile: string | null = null\nlet panicStatusBarItem: vscode.StatusBarItem | undefined\nlet panicButtonPressCount = 0\nlet panicButtonTimeoutId: NodeJS.Timeout | undefined\n\nconst CROWD_CODE_API_GATEWAY_URL = process.env.CROWD_CODE_API_GATEWAY_URL\nconst PANIC_BUTTON_TIMEOUT = 3000\nconst MAX_BUFFER_SIZE_PER_FILE = 1000 // Prevent unbounded growth\n\ninterface PendingEdit {\n\trangeOffset: number\n\trangeLength: number\n\ttext: string\n}\nconst pendingUserEdits = new Map<string, PendingEdit[]>()\n\n// Disposables for event subscriptions\nconst subscriptions: vscode.Disposable[] = []\n\n\nfunction logObservation(observation: Observation): void {\n\tif (!recording.isRecording) {return}\n\n\trecording.sequence++\n\tconst event: RecordingEvent = {\n\t\tsequence: recording.sequence,\n\t\ttimestamp: Date.now(),\n\t\ttype: 'observation',\n\t\tobservation,\n\t}\n\trecording.events.push(event)\n}\n\nfunction logAction(action: Action): void {\n\tif (!recording.isRecording) {return}\n\n\trecording.sequence++\n\tconst event: RecordingEvent = {\n\t\tsequence: recording.sequence,\n\t\ttimestamp: Date.now(),\n\t\ttype: 'action',\n\t\taction,\n\t}\n\trecording.events.push(event)\n}\n\n/**\n * Log an observation followed by an action (the standard pattern for user actions)\n */\nfunction logObservationAndAction(action: Action): void {\n\tlogObservation(captureObservation())\n\tlogAction(action)\n}\n\nexport function isCurrentFileExported(): boolean {\n    const editor = vscode.window.activeTextEditor\n    const filename = editor?.document.fileName.replaceAll('\\', '/')\n    const exportPath = getExportPath()\n    if (!editor || !filename || !exportPath) {\n        return false\n    }\n    return filename.startsWith(exportPath)\n}\n\n/**\n * Check if a change range is within the visible viewport\n * User edits must be within viewport; edits outside are from agents\n */\nfunction isChangeWithinViewport(\n\tchangeRange: vscode.Range,\n\tvisibleRanges: readonly vscode.Range[]\n): boolean {\n\treturn visibleRanges.some(visible =>\n\t\tvisible.contains(changeRange.start) || visible.contains(changeRange.end)\n\t)\n}\n\n/**\n * Apply user edits to content to reconstruct what the file would look like\n * if only the user had edited it (no agent changes)\n */\nfunction applyUserEdits(content: string, edits: PendingEdit[]): string {\n\t// Sort by offset descending to apply from end to start (preserves earlier offsets)\n\tconst sorted = [...edits].sort((a, b) => b.rangeOffset - a.rangeOffset)\n\n\tlet result = content\n\tfor (const edit of sorted) {\n\t\tresult = result.slice(0, edit.rangeOffset)\n\t\t\t+ edit.text\n\t\t\t+ result.slice(edit.rangeOffset + edit.rangeLength)\n\t}\n\treturn result\n}\n\n/**\n * Compute the agent-only diff by comparing user baseline to actual new content\n */\nfunction computeAgentOnlyDiff(\n\toldContent: string,\n\tnewContent: string,\n\tuserEdits: PendingEdit[],\n\tfilePath: string\n): string | null {\n\tconst userBaseline = applyUserEdits(oldContent, userEdits)\n\n\tif (userBaseline === newContent) {\n\t\treturn null // No agent changes\n\t}\n\n\tconst fileName = path.basename(filePath)\n\treturn createTwoFilesPatch(\n\t\t`a/${fileName}`,\n\t\t`b/${fileName}`,\n\t\tuserBaseline,\n\t\tnewContent,\n\t\t'',\n\t\t'',\n\t\t{ context: 3 }\n\t)\n}\n\nfunction handleTextDocumentChange(event: vscode.TextDocumentChangeEvent): void {\n\tif (!recording.isRecording) {return}\n\tif (isCurrentFileExported()) {return}\n\tif (event.document.uri.scheme !== 'file') {return}\n\n\tconst editor = vscode.window.activeTextEditor\n\n\t// Must be active document to be a user edit\n\tif (!editor || event.document !== editor.document) {return}\n\n\tconst visibleRanges = editor.visibleRanges\n\tconst file = vscode.workspace.asRelativePath(event.document.fileName)\n\n\tfor (const change of event.contentChanges) {\n\t\t// Drop changes outside viewport, these will be captured by filesystem watcher\n\t\tif (!isChangeWithinViewport(change.range, visibleRanges)) {\n\t\t\tcontinue\n\t\t}\n\n\t\t// This is a user edit, record it\n\t\tconst action: EditAction = {\n\t\t\tkind: 'edit',\n\t\t\tsource: 'user',\n\t\t\tfile,\n\t\t\tdiff: {\n\t\t\t\trangeOffset: change.rangeOffset,\n\t\t\t\trangeLength: change.rangeLength,\n\t\t\t\ttext: change.text,\n\t\t\t},\n\t\t}\n\n\t\tlogObservationAndAction(action)\n\n\t\t// Add to pending edits buffer for correlation with FS_CHANGE\n\t\tconst pendingEdit: PendingEdit = {\n\t\t\trangeOffset: change.rangeOffset,\n\t\t\trangeLength: change.rangeLength,\n\t\t\ttext: change.text,\n\t\t}\n\t\tconst edits = pendingUserEdits.get(file) ?? []\n\t\tif (edits.length < MAX_BUFFER_SIZE_PER_FILE) {\n\t\t\tedits.push(pendingEdit)\n\t\t\tpendingUserEdits.set(file, edits)\n\t\t}\n\t}\n\n\tactionsProvider.setCurrentFile(event.document.fileName)\n}\n\nfunction handleSelectionChange(event: vscode.TextEditorSelectionChangeEvent): void {\n\tif (!recording.isRecording) {return}\n\tif (event.textEditor !== vscode.window.activeTextEditor) {return}\n\tif (isCurrentFileExported()) {return}\n\n\tconst editor = event.textEditor\n\tconst selection = event.selections[0]\n\tif (!selection) {return}\n\n\tconst file = vscode.workspace.asRelativePath(editor.document.fileName)\n\tconst selectedText = editor.document.getText(selection)\n\n\tconst action: SelectionAction = {\n\t\tkind: 'selection',\n\t\tsource: 'user',\n\t\tfile,\n\t\tselectionStart: {\n\t\t\tline: selection.start.line,\n\t\t\tcharacter: selection.start.character,\n\t\t},\n\t\tselectionEnd: {\n\t\t\tline: selection.end.line,\n\t\t\tcharacter: selection.end.character,\n\t\t},\n\t\tselectedText,\n\t}\n\n\tlogObservationAndAction(action)\n\tactionsProvider.setCurrentFile(editor.document.fileName)\n}\n\nfunction handleActiveEditorChange(editor: vscode.TextEditor | undefined): void {\n\tupdateStatusBarItem()\n\t\n\tif (!recording.isRecording) {return}\n\tif (!editor) {return}\n\tif (isCurrentFileExported()) {return}\n\n\tconst file = vscode.workspace.asRelativePath(editor.document.fileName)\n\n\tconst action: TabSwitchAction = {\n\t\tkind: 'tab_switch',\n\t\tsource: 'user',\n\t\tfile,\n\t\tpreviousFile,\n\t}\n\n\tlogObservationAndAction(action)\n\t\n\tpreviousFile = file\n\tactionsProvider.setCurrentFile(editor.document.fileName)\n}\n\nfunction handleTerminalFocus(terminalId: string, terminalName: string): void {\n\tif (!recording.isRecording) {return}\n\tif (isCurrentFileExported()) {return}\n\n\tconst action: TerminalFocusAction = {\n\t\tkind: 'terminal_focus',\n\t\tsource: 'user',\n\t\tterminalId,\n\t\tterminalName,\n\t}\n\n\tlogObservationAndAction(action)\n\tactionsProvider.setCurrentFile(`Terminal: ${terminalName}`)\n}\n\nfunction handleTerminalCommand(terminalId: string, terminalName: string, command: string): void {\n\tif (!recording.isRecording) {return}\n\tif (isCurrentFileExported()) {return}\n\n\tconst action: TerminalCommandAction = {\n\t\tkind: 'terminal_command',\n\t\tsource: 'user',\n\t\tterminalId,\n\t\tterminalName,\n\t\tcommand,\n\t}\n\n\tlogObservationAndAction(action)\n}\n\nfunction handleTerminalOutput(terminalId: string, terminalName: string, output: string): void {\n\tif (!recording.isRecording) {return}\n\tif (isCurrentFileExported()) {return}\n\n\tconst action: TerminalOutputAction = {\n\t\tkind: 'terminal_output',\n\t\tsource: 'user',\n\t\tterminalId,\n\t\tterminalName,\n\t\toutput,\n\t}\n\n\t// Don't capture observation for every output chunk - just log the action\n\tlogAction(action)\n}\n\nexport function handleFileChange(\n\tfile: string,\n\tchangeType: 'create' | 'change' | 'delete',\n\toldContent: string | null,\n\tnewContent: string | null\n): void {\n\tif (!recording.isRecording) {return}\n\n\tconst relativePath = vscode.workspace.asRelativePath(file)\n\n\t// Helper to compute full diff\n\tconst computeFullDiff = (): string | null => {\n\t\tif (!oldContent && !newContent) {return null}\n\t\tif (oldContent === newContent) {return null}\n\t\tconst fileName = path.basename(file)\n\t\treturn createTwoFilesPatch(\n\t\t\t`a/${fileName}`,\n\t\t\t`b/${fileName}`,\n\t\t\toldContent ?? '',\n\t\t\tnewContent ?? '',\n\t\t\t'',\n\t\t\t'',\n\t\t\t{ context: 3 }\n\t\t)\n\t}\n\n\t// Check for git operation first\n\tconst gitOperation = getRecentGitOperation()\n\tif (gitOperation) {\n\t\tpendingUserEdits.clear() // Flush entire buffer, git can affect many files\n\t\tconst action: FileChangeAction = {\n\t\t\tkind: 'file_change',\n\t\t\tsource: gitOperation,\n\t\t\tfile: relativePath,\n\t\t\tchangeType,\n\t\t\tdiff: computeFullDiff(),\n\t\t}\n\t\tlogObservationAndAction(action)\n\t\treturn\n\t}\n\n\t// Get and clear pending user edits for this file\n\tconst pending = pendingUserEdits.get(relativePath)\n\tpendingUserEdits.delete(relativePath)\n\n\t// If no pending edits or missing content, record full diff as agent\n\tif (!pending || pending.length === 0 || !oldContent || !newContent) {\n\t\tconst action: FileChangeAction = {\n\t\t\tkind: 'file_change',\n\t\t\tsource: 'agent',\n\t\t\tfile: relativePath,\n\t\t\tchangeType,\n\t\t\tdiff: computeFullDiff(),\n\t\t}\n\t\tlogObservationAndAction(action)\n\t\treturn\n\t}\n\n\t// Three-way diff: compute agent-only changes\n\tconst agentDiff = computeAgentOnlyDiff(oldContent, newContent, pending, file)\n\n\t// Only record if there's remaining agent diff\n\tif (agentDiff) {\n\t\tconst action: FileChangeAction = {\n\t\t\tkind: 'file_change',\n\t\t\tsource: 'agent',\n\t\t\tfile: relativePath,\n\t\t\tchangeType,\n\t\t\tdiff: agentDiff,\n\t\t}\n\t\tlogObservationAndAction(action)\n\t}\n}\n\nfunction handleScrollObservation(observation: Observation): void {\n\tif (!recording.isRecording) {return}\n\t\n\tconst editor = vscode.window.activeTextEditor\n\tif (!editor) {return}\n\n\tconst file = vscode.workspace.asRelativePath(editor.document.fileName)\n\n\tlogObservation(observation)\n\n\tconst action: ScrollAction = {\n\t\tkind: 'scroll',\n\t\tsource: 'user',\n\t\tfile,\n\t}\n\tlogAction(action)\n}\n\nfunction createRecordingFolder(folderPath: string): void {\n    if (!fs.existsSync(folderPath)) {\n        fs.mkdirSync(folderPath, { recursive: true })\n    }\n}\n\nexport async function startRecording(): Promise<void> {\n    if (recording.isRecording) {\n        notificationWithProgress('Already recording')\n        logToOutput('Already recording', 'info')\n        return\n    }\n\n    const exportPath = getExportPath()\n    if (!exportPath) {\n        return\n    }\n\n\t// Add to gitignore if configured\n    if (\n        getConfig().get<boolean>('export.addToGitignore') &&\n        getConfig().get<string>('export.exportPath')?.startsWith('${workspaceFolder}')\n    ) {\n        await addToGitignore()\n    }\n\n\t// Initialize recording state\n    recording.startDateTime = new Date()\n\trecording.endDateTime = null\n\trecording.sequence = 0\n\trecording.events = []\n\trecording.sessionId = vscode.env.sessionId\n\tpreviousFile = null\n\tpanicButtonPressCount = 0\n\ttimer = 0\n\n\t// Reset capture module states\n\tresetObservationState()\n\tresetTerminalState()\n\tresetFilesystemState()\n\tpendingUserEdits.clear()\n\n\t// Create recording folder\n\tconst baseFilePath = generateBaseFilePath(recording.startDateTime, false, undefined, recording.sessionId)\n    if (!baseFilePath) {\n        return\n    }\n    const folderPath = path.dirname(path.join(exportPath, baseFilePath))\n    createRecordingFolder(folderPath)\n\n\t// Initialize capture modules with callbacks\n\tinitializeViewportCapture(extContext, handleScrollObservation)\n\tinitializeTerminalCapture(extContext, {\n\t\tonFocus: handleTerminalFocus,\n\t\tonCommand: handleTerminalCommand,\n\t\tonOutput: handleTerminalOutput,\n\t})\n\tawait initializeFilesystemWatcher(extContext, handleFileChange)\n\n\t// Subscribe to VS Code events\n\tsubscriptions.push(\n\t\tvscode.workspace.onDidChangeTextDocument(handleTextDocumentChange)\n\t)\n\tsubscriptions.push(\n\t\tvscode.window.onDidChangeTextEditorSelection(handleSelectionChange)\n\t)\n\tsubscriptions.push(\n\t\tvscode.window.onDidChangeActiveTextEditor(handleActiveEditorChange)\n\t)\n\n    recording.isRecording = true\n\n\t// Start timer\n    intervalId = setInterval(() => {\n\t\ttimer++\n        updateStatusBarItem()\n    }, 1000)\n\n\t// Capture initial observation\n\tconst initialObservation = captureObservation()\n\tlogObservation(initialObservation)\n\n\t// Set up upload interval\n\tuploadIntervalId = setInterval(async () => {\n\t\tawait uploadRecording()\n\t}, 5 * 60 * 1000) // 5 minutes\n\n    notificationWithProgress('Recording started')\n\tlogToOutput('Recording started (v2.0)', 'info')\n\n    updateStatusBarItem()\n    updatePanicButton()\n    actionsProvider.setRecordingState(true)\n\n\t// Set current file\n\tconst editor = vscode.window.activeTextEditor\n\tif (editor) {\n\t\tpreviousFile = vscode.workspace.asRelativePath(editor.document.fileName)\n\t\tactionsProvider.setCurrentFile(editor.document.fileName)\n                }\n            }\n\nexport async function stopRecording(force = false): Promise<void> {\n    if (!recording.isRecording) {\n        notificationWithProgress('Not recording')\n        return\n    }\n\n    recording.isRecording = false\n\trecording.endDateTime = new Date()\n\n\t// Clear intervals\n\tif (intervalId) {\n    clearInterval(intervalId)\n\t\tintervalId = null\n\t}\n\tif (uploadIntervalId) {\n\t\tclearInterval(uploadIntervalId)\n\t\tuploadIntervalId = null\n\t}\n    if (panicButtonTimeoutId) {\n        clearTimeout(panicButtonTimeoutId)\n        panicButtonTimeoutId = undefined\n    }\n\n\t// Dispose subscriptions\n\tfor (const subscription of subscriptions) {\n\t\tsubscription.dispose()\n\t}\n\tsubscriptions.length = 0\n\n\ttimer = 0\n\tpanicButtonPressCount = 0\n\n    updateStatusBarItem()\n    updatePanicButton()\n    actionsProvider.setRecordingState(false)\n\n    if (force) {\n        notificationWithProgress('Recording cancelled')\n        logToOutput('Recording cancelled', 'info')\n\t\trecording.events = []\n        return\n    }\n\n\t// Save recording\n\tawait saveRecording()\n\n    notificationWithProgress('Recording finished')\n\tlogToOutput('Recording finished (v2.0)', 'info')\n}\n\n\nasync function saveRecording(): Promise<void> {\n\tconst exportPath = getExportPath()\n\tif (!exportPath || !recording.startDateTime) {\n        return\n    }\n\n\tconst baseFilePath = generateBaseFilePath(recording.startDateTime, false, undefined, recording.sessionId)\n\tif (!baseFilePath) {\n        return\n    }\n\n\tconst session: RecordingSession = {\n\t\tversion: '2.0',\n\t\tsessionId: recording.sessionId,\n\t\tstartTime: recording.startDateTime.getTime(),\n\t\tevents: recording.events,\n\t}\n\n\tconst jsonContent = JSON.stringify(session, null, 2)\n\tconst filePath = path.join(exportPath, `${baseFilePath}.json`)\n\n        try {\n            const directory = path.dirname(filePath)\n            if (!fs.existsSync(directory)) {\n                fs.mkdirSync(directory, { recursive: true })\n            }\n\t\tawait fs.promises.writeFile(filePath, jsonContent)\n\t\tlogToOutput(`Recording saved to ${filePath}`, 'info')\n        } catch (err) {\n\t\tlogToOutput(`Failed to save recording: ${err}`, 'error')\n}\n\n\t// Refresh the recordFiles view\n\tvscode.commands.executeCommand('crowd-code.refreshRecordFiles')\n}\n\nasync function uploadRecording(): Promise<void> {\n\tif (!recording.isRecording) {return}\n\tif (!hasConsent()) {return}\n\tif (typeof CROWD_CODE_API_GATEWAY_URL !== 'string' || !CROWD_CODE_API_GATEWAY_URL.trim()) {\n        return\n    }\n\n    const exportPath = getExportPath()\n\tif (!exportPath || !recording.startDateTime) {\n        return\n    }\n\n\tconst baseFilePath = generateBaseFilePath(recording.startDateTime, false, undefined, recording.sessionId)\n\tif (!baseFilePath) {\n        return\n    }\n\n\tconst session: RecordingSession = {\n\t\tversion: '2.0',\n\t\tsessionId: recording.sessionId,\n\t\tstartTime: recording.startDateTime.getTime(),\n\t\tevents: recording.events,\n\t}\n\n\tconst jsonContent = JSON.stringify(session)\n\tconst extensionVersion = extContext.extension.packageJSON.version as string\n\tconst userId = extContext.globalState.get<string>('userId')\n\n\ttry {\n\t\tconst payload = {\n\t\t\tfileName: `${baseFilePath}.json`,\n\t\t\tcontent: jsonContent,\n\t\t\tversion: extensionVersion,\n\t\t\tuserId,\n\t\t}\n\t\tawait axios.post(CROWD_CODE_API_GATEWAY_URL, payload)\n\t\tlogToOutput(`Successfully uploaded recording`, 'info')\n\t} catch (error: unknown) {\n\t\tif (axios.isAxiosError(error)) {\n\t\t\tlogToOutput(`Error uploading recording: ${error.message}`, 'error')\n\t\t}\n\t}\n}\n\n\nexport function updateStatusBarItem(): void {\n    if (recording.isRecording) {\n        if (getConfig().get('appearance.showTimer') === false) {\n            statusBarItem.text = '$(debug-stop)'\n\t\t\tstatusBarItem.tooltip = 'Current time: ' + formatDisplayTime(timer)\n\t\t} else {\n\t\t\tstatusBarItem.text = '$(debug-stop) ' + formatDisplayTime(timer)\n            statusBarItem.tooltip = 'Stop Recording'\n        }\n        statusBarItem.command = commands.stopRecording\n        statusBarItem.show()\n    } else {\n        const editor = vscode.window.activeTextEditor\n        if (!editor) {\n            statusBarItem.hide()\n            return\n        }\n        if (getConfig().get('appearance.minimalMode') === true) {\n            statusBarItem.text = '$(circle-large-filled)'\n        } else {\n            statusBarItem.text = '$(circle-large-filled) Start Recording'\n        }\n        statusBarItem.tooltip = 'Start Recording'\n        statusBarItem.command = commands.startRecording\n        statusBarItem.show()\n    }\n}\n\n\nexport function updatePanicButton(): void {\n    if (!recording.isRecording) {\n        if (panicStatusBarItem) {\n            panicStatusBarItem.hide()\n        }\n        return\n    }\n\n    if (!panicStatusBarItem) {\n\t\tpanicStatusBarItem = vscode.window.createStatusBarItem(vscode.StatusBarAlignment.Right, 8999)\n        extContext.subscriptions.push(panicStatusBarItem)\n    }\n\n\tconst secondsToRemove = (panicButtonPressCount + 1) * 10\n    panicStatusBarItem.text = '$(refresh)'\n\tpanicStatusBarItem.tooltip = `Remove last ${secondsToRemove} seconds of recording`\n    panicStatusBarItem.command = commands.panicButton\n    panicStatusBarItem.show()\n}\n\nexport async function panicButton(): Promise<void> {\n    if (!recording.isRecording) {\n        vscode.window.showWarningMessage('No active recording to remove data from')\n        return\n    }\n\n    if (!recording.startDateTime) {\n        vscode.window.showErrorMessage('Recording start time not available')\n        return\n    }\n\n\tconst secondsToRemove = (panicButtonPressCount + 1) * 10\n\tconst cutoffTime = Date.now() - (secondsToRemove * 1000)\n\n\t// Remove events after cutoff time\n\tconst originalCount = recording.events.length\n\trecording.events = recording.events.filter(event => event.timestamp < cutoffTime)\n\tconst removedCount = originalCount - recording.events.length\n\n\t// Update sequence to match\n\tif (recording.events.length > 0) {\n\t\trecording.sequence = recording.events[recording.events.length - 1].sequence\n\t} else {\n\t\trecording.sequence = 0\n\t}\n\n        panicButtonPressCount++\n        \n\t// Reset timeout\n        if (panicButtonTimeoutId) {\n            clearTimeout(panicButtonTimeoutId)\n        }\n        panicButtonTimeoutId = setTimeout(() => {\n            panicButtonPressCount = 0\n            updatePanicButton()\n        }, PANIC_BUTTON_TIMEOUT)\n        \n        updatePanicButton()\n                \n                vscode.window.showInformationMessage(\n\t\t`Removed ${removedCount} events (last ${secondsToRemove} seconds)`,\n                    'Dismiss'\n                )\n            }\n",typescript,tab
+2,478,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"9:16:35 AM [info] Activating crowd-code\n9:16:35 AM [info] Recording started\n9:16:35 AM [info] Initializing git provider using file system watchers...\n9:16:35 AM [info] Git repository found\n9:16:35 AM [info] Git provider initialized successfully\n9:16:35 AM [info] Initial git state: [object Object]\n",Log,tab
+3,1384,"src/recording.ts",0,0,"",typescript,tab
+4,4129,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+5,4253,"extension-output-pdoom-org.crowd-code-#1-crowd-code",37,0,"",Log,selection_command
+6,4842,"extension-output-pdoom-org.crowd-code-#1-crowd-code",40,0,"",Log,selection_command
+7,6440,"src/recording.ts",0,0,"",typescript,tab
+8,6442,"TERMINAL",0,0,"",,terminal_focus
+9,8371,"TERMINAL",0,0,"squeue",,terminal_command
+10,8381,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  57887 franz.sram interacti     1    20  R 2026-01-08T17:20:08 2026-01-08T17:20:08 15:56:35 1-00:00:00 hai001\r\n  57882 franz.sram interacti     1    20  R 2026-01-08T15:42:55 2026-01-08T15:42:55 17:33:48 1-00:00:00 hai001\r\n  57998 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-09T01:09:29  8:07:14 1-00:00:00 hai008\r\n  57997 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-09T01:09:28  8:07:15 1-00:00:00 hai008\r\n  57996 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-09T00:00:43  9:16:00 1-00:00:00 hai006\r\n  57995 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-08T23:52:34  9:24:09 1-00:00:00 hai002\r\n  57993 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-08T22:30:06 10:46:37 1-00:00:00 hai006\r\n  57994 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-08T22:30:06 10:46:37 1-00:00:00 hai007\r\n  57989 yannick.ki  standard     1    32  R 2026-01-08T22:29:39 2026-01-08T22:29:40 10:47:03 1-00:00:00 hai002\r\n  57990 yannick.ki  standard     1    32  R 2026-01-08T22:29:39 2026-01-08T22:29:40 10:47:03 1-00:00:00 hai006\r\n  57991 yannick.ki  standard     1    32  R 2026-01-08T22:29:39 2026-01-08T22:29:40 10:47:03 1-00:00:00 hai007\r\n  57883 nishant.ku  standard     3   624  R 2026-01-08T16:50:21 2026-01-08T16:50:22 16:26:21 1-00:00:00 hai[003-005]\r\n  57867 franz.sram  standard     1    16  R 2026-01-08T11:52:19 2026-01-08T11:52:19 21:24:24 1-00:00:00 hai001\r\n]0;franz.srambical@hai-login1:~/crowd-code",,terminal_output
+11,26561,"TERMINAL",0,0,"scancel 57887 57882",,terminal_command
+12,26569,"TERMINAL",0,0,"]633;C]0;franz.srambical@hai-login1:~/crowd-code",,terminal_output
+13,31169,"TERMINAL",0,0,"squeue",,terminal_command
+14,31171,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  57887 franz.sram interacti     1    20 CG 2026-01-08T17:20:08 2026-01-08T17:20:08 15:56:53 1-00:00:00 hai001\r\n  57882 franz.sram interacti     1    20 CG 2026-01-08T15:42:55 2026-01-08T15:42:55 17:34:06 1-00:00:00 hai001\r\n  57998 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-09T01:09:29  8:07:37 1-00:00:00 hai008\r\n  57997 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-09T01:09:28  8:07:38 1-00:00:00 hai008\r\n  57996 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-09T00:00:43  9:16:23 1-00:00:00 hai006\r\n  57995 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-08T23:52:34  9:24:32 1-00:00:00 hai002\r\n  57993 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-08T22:30:06 10:47:00 1-00:00:00 hai006\r\n  57994 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-08T22:30:06 10:47:00 1-00:00:00 hai007\r\n  57989 yannick.ki  standard     1    32  R 2026-01-08T22:29:39 2026-01-08T22:29:40 10:47:26 1-00:00:00 hai002\r\n  57990 yannick.ki  standard     1    32  R 2026-01-08T22:29:39 2026-01-08T22:29:40 10:47:26 1-00:00:00 hai006\r\n  57991 yannick.ki  standard     1    32  R 2026-01-08T22:29:39 2026-01-08T22:29:40 10:47:26 1-00:00:00 hai007\r\n  57883 nishant.ku  standard     3   624  R 2026-01-08T16:50:21 2026-01-08T16:50:22 16:26:44 1-00:00:00 hai[003-005]\r\n  57867 franz.sram  standard     1    16  R 2026-01-08T11:52:19 2026-01-08T11:52:19 21:24:47 1-00:00:00 hai001\r\n]0;franz.srambical@hai-login1:~/crowd-code",,terminal_output
+15,1008018,"src/recording.ts",4733,0,"",typescript,selection_mouse
+16,1008020,"src/recording.ts",4732,0,"",typescript,selection_command
+17,1010582,"src/gitProvider.ts",0,0,"/**\n * Git Provider for crowd-code 2.0\n * Detects git operations to annotate filesystem changes\n */\n\nimport * as vscode from 'vscode'\nimport { logToOutput } from './utilities'\n\n// Track recent git operations for filesystem change attribution\nlet sawHeadChange = false\nlet lastGitOperationTime = 0\nconst GIT_OPERATION_WINDOW_MS = 5000\n\n// File system watchers\nlet gitHeadWatcher: vscode.FileSystemWatcher | undefined\nlet gitRefsWatcher: vscode.FileSystemWatcher | undefined\n\n/**\n * Check if there was a recent git operation\n * Returns 'git_checkout' if HEAD changed, 'git' for other operations, null otherwise\n */\nexport function getRecentGitOperation(): 'git' | 'git_checkout' | null {\n\tif (Date.now() - lastGitOperationTime > GIT_OPERATION_WINDOW_MS) {\n\t\treturn null\n\t}\n\tconst result = sawHeadChange ? 'git_checkout' : 'git'\n\tsawHeadChange = false\n\treturn result\n}\n\n/**\n * Setup git file watchers for the current workspace\n */\nfunction setupGitWatchers(): void {\n\t// Cleanup any existing watchers first\n\tdisposeWatchers()\n\n\tconst workspaceFolder = vscode.workspace.workspaceFolders?.[0]\n\tif (!workspaceFolder) {\n\t\tlogToOutput('No workspace folder found', 'info')\n\t\treturn\n\t}\n\n\tconst gitDir = vscode.Uri.joinPath(workspaceFolder.uri, '.git')\n\tvscode.workspace.fs.stat(gitDir).then(\n\t\t() => {\n\t\t\tlogToOutput('Git repository found', 'info')\n\n\t\t\t// Watch .git/HEAD for branch changes\n\t\t\tgitHeadWatcher = vscode.workspace.createFileSystemWatcher(\n\t\t\t\tnew vscode.RelativePattern(workspaceFolder, '.git/HEAD')\n\t\t\t)\n\t\t\tgitHeadWatcher.onDidChange(() => {\n\t\t\t\tlogToOutput('Git checkout detected', 'info')\n\t\t\t\tsawHeadChange = true\n\t\t\t\tlastGitOperationTime = Date.now()\n\t\t\t})\n\n\t\t\t// Watch .git/refs for other git operations (pull, stash, etc.)\n\t\t\tgitRefsWatcher = vscode.workspace.createFileSystemWatcher(\n\t\t\t\tnew vscode.RelativePattern(workspaceFolder, '.git/refs/**/*')\n\t\t\t)\n\t\t\tgitRefsWatcher.onDidChange(() => {\n\t\t\t\tlogToOutput('Git refs changed', 'info')\n\t\t\t\tlastGitOperationTime = Date.now()\n\t\t\t})\n\t\t\tgitRefsWatcher.onDidCreate(() => {\n\t\t\t\tlogToOutput('Git refs created', 'info')\n\t\t\t\tlastGitOperationTime = Date.now()\n\t\t\t})\n\t\t\tgitRefsWatcher.onDidDelete(() => {\n\t\t\t\tlogToOutput('Git refs deleted', 'info')\n\t\t\t\tlastGitOperationTime = Date.now()\n\t\t\t})\n\n\t\t\tlogToOutput('Git provider initialized', 'info')\n\t\t},\n\t\t() => {\n\t\t\tlogToOutput('Not a git repository', 'info')\n\t\t}\n\t)\n}\n\nfunction disposeWatchers(): void {\n\tgitHeadWatcher?.dispose()\n\tgitHeadWatcher = undefined\n\tgitRefsWatcher?.dispose()\n\tgitRefsWatcher = undefined\n}\n\n/**\n * Initialize the git provider\n */\nexport function initializeGitProvider(context: vscode.ExtensionContext): void {\n\tlogToOutput('Initializing git provider...', 'info')\n\n\tsetupGitWatchers()\n\n\t// Reinitialize on workspace changes\n\tcontext.subscriptions.push(\n\t\tvscode.workspace.onDidChangeWorkspaceFolders(() => {\n\t\t\tlogToOutput('Workspace changed, reinitializing git provider...', 'info')\n\t\t\tsetupGitWatchers()\n\t\t})\n\t)\n}\n\n/**\n * Cleanup the git provider\n */\nexport function cleanupGitProvider(): void {\n\tdisposeWatchers()\n\tsawHeadChange = false\n\tlastGitOperationTime = 0\n}\n",typescript,tab
+18,1011288,"src/gitProvider.ts",0,0,"",typescript,selection_command
+19,1011732,"src/gitProvider.ts",4,0,"",typescript,selection_command
+20,1011984,"src/gitProvider.ts",39,0,"",typescript,selection_command
+21,1012013,"src/gitProvider.ts",96,0,"",typescript,selection_command
+22,1012057,"src/gitProvider.ts",100,0,"",typescript,selection_command
+23,1012081,"src/gitProvider.ts",101,0,"",typescript,selection_command
+24,1012113,"src/gitProvider.ts",134,0,"",typescript,selection_command
+25,1012147,"src/gitProvider.ts",176,0,"",typescript,selection_command
+26,1012376,"src/gitProvider.ts",177,0,"",typescript,selection_command
+27,1012377,"src/gitProvider.ts",242,0,"",typescript,selection_command
+28,1012378,"src/gitProvider.ts",268,0,"",typescript,selection_command
+29,1012378,"src/gitProvider.ts",297,0,"",typescript,selection_command
+30,1012867,"src/gitProvider.ts",297,36,"const GIT_OPERATION_WINDOW_MS = 5000",typescript,selection_command
+31,1123927,"src/gitProvider.ts",415,0,"",typescript,selection_mouse
+32,1123932,"src/gitProvider.ts",414,0,"",typescript,selection_command
+33,1124456,"src/gitProvider.ts",334,0,"",typescript,selection_mouse
+34,1528649,"src/gitProvider.ts",0,0,"",typescript,tab
+35,1528651,"src/gitProvider.ts",297,0,"",typescript,selection_command
+36,1569868,"src/gitProvider.ts",0,0,"",typescript,tab
+37,2315546,"src/gitProvider.ts",333,0,"\n",typescript,content
+38,2316648,"src/gitProvider.ts",333,1,"",typescript,content
+39,2318293,"src/gitProvider.ts",333,0,"\n",typescript,content
+40,2319206,"src/gitProvider.ts",333,1,"",typescript,content
+41,2323966,"src/recording.ts",0,0,"",typescript,tab
+42,2937704,"src/gitProvider.ts",0,0,"",typescript,tab
+43,3023326,"src/gitProvider.ts",0,0,"",typescript,tab
+44,3023327,"src/gitProvider.ts",297,0,"",typescript,selection_command
+45,3036832,"src/gitProvider.ts",0,0,"",typescript,tab
+46,3038075,"src/recording.ts",0,0,"",typescript,tab
+47,3576568,"src/test/recording.test.ts",0,0,"import * as assert from 'node:assert'\r\nimport * as vscode from 'vscode'\r\nimport * as path from 'node:path'\r\nimport * as fs from 'node:fs'\r\nimport { setDefaultOptions, getConfig } from '../utilities'\r\nimport { statusBarItem } from '../extension'\r\n\r\n/**\r\n * Waits for the specified number of milliseconds and then resolves the returned Promise.\r\n * @param ms - The number of milliseconds to wait. Defaults to 500 ms.\r\n * @returns A Promise that resolves after the specified number of milliseconds.\r\n */\r\nconst waitMs = (ms = 500) => new Promise(resolve => setTimeout(resolve, ms))\r\n\r\nsuite('Recording Tests', () => {\r\n\tconst publisher = 'pdoom-org'\r\n\tconst extensionName = 'crowd-code'\r\n\r\n\tlet workspaceFolder: string\r\n\r\n\tlet statusBarSpy: vscode.StatusBarItem\r\n\r\n\tvscode.window.showInformationMessage('Start all tests.')\r\n\r\n\tsetup(async () => {\r\n\t\t// biome-ignore lint/style/noNonNullAssertion: the workspace folder is created by the test suite\r\n\t\tworkspaceFolder = vscode.workspace.workspaceFolders![0].uri.fsPath\r\n\t\tsetDefaultOptions()\r\n\t\t// set workspace export path\r\n\t\tgetConfig().update(\r\n\t\t\t'export.exportPath',\r\n\t\t\t'${workspaceFolder}',\r\n\t\t\tvscode.ConfigurationTarget.Workspace\r\n\t\t)\r\n\t\tstatusBarSpy = statusBarItem\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.stopRecording`)\r\n\t})\r\n\r\n\tteardown(async () => {\r\n\t\t// First ensure recording is stopped\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.stopRecording`)\r\n\r\n\t\t// Add small delay to ensure VS Code releases file handles\r\n\t\tawait vscode.commands.executeCommand('workbench.action.closeAllEditors')\r\n\t\tawait waitMs(100)\r\n\r\n\t\tif (workspaceFolder) {\r\n\t\t\tconst files = fs.readdirSync(workspaceFolder)\r\n\t\t\tfor (const file of files) {\r\n\t\t\t\tif (file !== '.vscode') {\r\n\t\t\t\t\tfs.unlinkSync(path.join(workspaceFolder, file))\r\n\t\t\t\t}\r\n\t\t\t}\r\n\t\t}\r\n\t})\r\n\r\n\ttest('Should be visible the status bar item', async () => {\r\n\t\t// Wait for status bar item to be created\r\n\t\tawait waitMs()\r\n\r\n\t\tassert.strictEqual(\r\n\t\t\tstatusBarSpy.text.includes('$(circle-large-filled)'),\r\n\t\t\ttrue,\r\n\t\t\t'Should be visible the circle icon'\r\n\t\t)\r\n\t\tassert.strictEqual(\r\n\t\t\tstatusBarSpy.tooltip?.toString().includes('Start Recording'),\r\n\t\t\ttrue,\r\n\t\t\t'The tooltip should be ""Start Recording""'\r\n\t\t)\r\n\t})\r\n\r\n\ttest('Should start recording when start command is executed', async () => {\r\n\t\t// Execute start recording command\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.startRecording`)\r\n\r\n\t\t// Get status bar state through the extension's status bar item\r\n\t\tassert.strictEqual(\r\n\t\t\tstatusBarSpy.text.includes('$(debug-stop)'),\r\n\t\t\ttrue,\r\n\t\t\t'Should be visible the stop icon'\r\n\t\t)\r\n\t\tassert.strictEqual(\r\n\t\t\tstatusBarSpy.tooltip?.toString().includes('Stop Recording'),\r\n\t\t\ttrue,\r\n\t\t\t""Status bar item tooltip should be 'Stop Recording'""\r\n\t\t)\r\n\t})\r\n\r\n\ttest('Should create CSV file when recording starts', async () => {\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.startRecording`)\r\n\r\n\t\t// Wait for file creation\r\n\t\tawait waitMs(1000)\r\n\r\n\t\tconst files = fs.readdirSync(workspaceFolder)\r\n\t\tconst csvFile = files.find(file => file.endsWith('.csv'))\r\n\r\n\t\tassert.ok(csvFile, 'CSV file should be created')\r\n\r\n\t\t// Cleanup\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.stopRecording`)\r\n\t})\r\n\r\n\ttest('Should stop recording when stop command is executed', async () => {\r\n\t\t// Start recording first\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.startRecording`)\r\n\t\tawait waitMs(1000)\r\n\r\n\t\t// Stop recording\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.stopRecording`)\r\n\r\n\t\t// Check status bar\r\n\t\tassert.strictEqual(statusBarSpy.text.includes('$(circle-large-filled)'), true)\r\n\t\tassert.strictEqual(statusBarSpy.tooltip?.toString().includes('Start Recording'), true)\r\n\t})\r\n\r\n\ttest('Should not generate output files when stopping recording', async () => {\r\n\t\t// Configure export formats (none)\r\n\t\tawait getConfig().update('exportFormats', [])\r\n\r\n\t\t// Start and stop recording\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.startRecording`)\r\n\t\tawait waitMs(1000)\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.stopRecording`)\r\n\t\tawait waitMs()\r\n\r\n\t\t// Check for output files\r\n\t\tconst files = fs.readdirSync(workspaceFolder)\r\n\t\tconst jsonFile = files.find(file => file.endsWith('.json'))\r\n\t\tconst srtFile = files.find(file => file.endsWith('.srt'))\r\n\r\n\t\tassert.ok(!jsonFile, 'JSON file should NOT be created')\r\n\t\tassert.ok(!srtFile, 'SRT file should NOT be created')\r\n\t})\r\n\r\n\ttest('Should generate JSON output file when stopping recording', async () => {\r\n\t\t// Configure export formats\r\n\t\tawait getConfig().update('exportFormats', ['JSON'])\r\n\r\n\t\t// Start and stop recording\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.startRecording`)\r\n\t\tawait waitMs(1000)\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.stopRecording`)\r\n\t\tawait waitMs()\r\n\r\n\t\t// Check for output files\r\n\t\tconst files = fs.readdirSync(workspaceFolder)\r\n\t\tconst jsonFile = files.find(file => file.endsWith('.json'))\r\n\t\tconst srtFile = files.find(file => file.endsWith('.srt'))\r\n\r\n\t\tassert.ok(jsonFile, 'JSON file should be created')\r\n\t\tassert.ok(!srtFile, 'SRT file should NOT be created')\r\n\t})\r\n\r\n\ttest('Should generate SRT output file when stopping recording', async () => {\r\n\t\t// Configure export formats\r\n\t\tawait getConfig().update('exportFormats', ['SRT'])\r\n\r\n\t\t// Start and stop recording\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.startRecording`)\r\n\t\tawait waitMs(1000)\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.stopRecording`)\r\n\t\tawait waitMs()\r\n\r\n\t\t// Check for output files\r\n\t\tconst files = fs.readdirSync(workspaceFolder)\r\n\t\tconst jsonFile = files.find(file => file.endsWith('.json'))\r\n\t\tconst srtFile = files.find(file => file.endsWith('.srt'))\r\n\r\n\t\tassert.ok(!jsonFile, 'JSON file should NOT be created')\r\n\t\tassert.ok(srtFile, 'SRT file should be created')\r\n\t})\r\n\r\n\ttest('Should generate output files (JSON, SRT) when stopping recording', async () => {\r\n\t\t// Configure export formats\r\n\t\tawait getConfig().update('exportFormats', ['JSON', 'SRT'])\r\n\r\n\t\t// Start and stop recording\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.startRecording`)\r\n\t\tawait waitMs(1000)\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.stopRecording`)\r\n\t\tawait waitMs()\r\n\r\n\t\t// Check for output files\r\n\t\tconst files = fs.readdirSync(workspaceFolder)\r\n\t\tconst jsonFile = files.find(file => file.endsWith('.json'))\r\n\t\tconst srtFile = files.find(file => file.endsWith('.srt'))\r\n\r\n\t\tassert.ok(jsonFile, 'JSON file should be created')\r\n\t\tassert.ok(srtFile, 'SRT file should be created')\r\n\t})\r\n\r\n\tconst testCsvFile = (csvPath: string, expectedLines: string[]) => {\r\n\t\tconst csvContent = fs.readFileSync(csvPath, 'utf-8')\r\n\t\tconst lines = csvContent.split('\n').filter(line => line.trim() !== '')\r\n\r\n\t\tassert.strictEqual(\r\n\t\t\tlines.length,\r\n\t\t\texpectedLines.length,\r\n\t\t\t'Number of lines in CSV file should match expected lines'\r\n\t\t)\r\n\r\n\t\tfor (let i = 0; i < lines.length; i++) {\r\n\t\t\tconst line = lines[i]\r\n\t\t\tlet expectedLine = expectedLines[i]\r\n\r\n\t\t\tconst timestampIndex = expectedLine.indexOf('%n')\r\n\t\t\tif (timestampIndex !== -1) {\r\n\t\t\t\tconst commaIndex = expectedLine.indexOf(',', timestampIndex + 1)\r\n\t\t\t\texpectedLine = expectedLine.replace('%n', line.substring(0, commaIndex))\r\n\t\t\t}\r\n\t\t\tassert.strictEqual(\r\n\t\t\t\tlines[i],\r\n\t\t\t\texpectedLines[i],\r\n\t\t\t\t`Line ${i + 1} in CSV file should match expected line`\r\n\t\t\t)\r\n\t\t}\r\n\t}\r\n\r\n\ttest('Should record file changes and verify exports', async () => {\r\n\t\t// Create and write to new file using VS Code API\r\n\t\tconst testFileUri = vscode.Uri.file(path.join(workspaceFolder, 'test.txt'))\r\n\t\tconst initialContent = 'This is an example recording'\r\n\t\tawait vscode.workspace.fs.writeFile(testFileUri, Buffer.from(''))\r\n\r\n\t\t// Open file in VS Code\r\n\t\tconst doc = await vscode.workspace.openTextDocument(testFileUri)\r\n\t\tconst editor = await vscode.window.showTextDocument(doc)\r\n\r\n\t\t// Start recording\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.startRecording`)\r\n\t\tawait waitMs()\r\n\r\n\t\t// Get CSV path\r\n\t\tconst csvFilename = fs.readdirSync(workspaceFolder).find(f => f.endsWith('.csv'))\r\n\r\n\t\tassert.strictEqual(csvFilename !== undefined, true, 'CSV file should be created')\r\n\r\n\t\tif (csvFilename === undefined) {\r\n\t\t\treturn\r\n\t\t}\r\n\t\tconst csvPath = path.join(workspaceFolder, csvFilename)\r\n\r\n\t\tconst csvExpectedLines = [\r\n\t\t\t'Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type',\r\n\t\t\t'1,%n,""test.txt"",0,0,"""",plaintext,tab',\r\n\t\t]\r\n\t\ttestCsvFile(csvPath, csvExpectedLines)\r\n\r\n\t\tawait editor.edit(editBuilder => {\r\n\t\t\teditBuilder.insert(new vscode.Position(0, 0), initialContent)\r\n\t\t})\r\n\t\tawait waitMs(1000)\r\n\r\n\t\tcsvExpectedLines.push('2,%n,""test.txt"",0,0,""This is an example recording"",plaintext,content')\r\n\t\ttestCsvFile(csvPath, csvExpectedLines)\r\n\r\n\t\t// Select text to remove\r\n\t\tconst textToRemove = 'n example'\r\n\t\tconst startPos = initialContent.indexOf(textToRemove)\r\n\t\tawait editor.edit(editBuilder => {\r\n\t\t\teditBuilder.replace(\r\n\t\t\t\tnew vscode.Range(\r\n\t\t\t\t\tnew vscode.Position(0, startPos),\r\n\t\t\t\t\tnew vscode.Position(0, startPos + textToRemove.length)\r\n\t\t\t\t),\r\n\t\t\t\t''\r\n\t\t\t)\r\n\t\t})\r\n\t\tawait waitMs(1000)\r\n\r\n\t\tcsvExpectedLines.push('3,%n,""test.txt"",0,0,""This is a recording"",plaintext,content')\r\n\t\ttestCsvFile(csvPath, csvExpectedLines)\r\n\r\n\t\t// Stop recording and wait for export\r\n\t\tawait vscode.commands.executeCommand(`${extensionName}.stopRecording`)\r\n\t\tawait waitMs(1000)\r\n\r\n\t\t// Verify exports\r\n\t\tconst files = fs.readdirSync(workspaceFolder)\r\n\t\tconst jsonFile = files.find(f => f.endsWith('.json'))\r\n\t\tconst srtFile = files.find(f => f.endsWith('.srt'))\r\n\r\n\t\tif (jsonFile) {\r\n\t\t\tconst jsonContent = JSON.parse(fs.readFileSync(path.join(workspaceFolder, jsonFile), 'utf-8'))\r\n\t\t\t// biome-ignore lint/suspicious/noExplicitAny: <explanation>\r\n\t\t\tassert.ok(jsonContent.some((change: any) => change.text.includes(initialContent)))\r\n\t\t\t// biome-ignore lint/suspicious/noExplicitAny: <explanation>\r\n\t\t\tassert.ok(jsonContent.some((change: any) => change.text.includes('This is a recording')))\r\n\t\t}\r\n\r\n\t\tif (srtFile) {\r\n\t\t\tconst srtContent = fs.readFileSync(path.join(workspaceFolder, srtFile), 'utf-8')\r\n\t\t\tassert.ok(srtContent.includes(initialContent))\r\n\t\t\tassert.ok(srtContent.includes('This is a recording'))\r\n\t\t}\r\n\t})\r\n})\r\n",typescript,tab
+48,3584552,"src/recording.ts",0,0,"",typescript,tab
+49,3603024,"src/extension.ts",0,0,"/**\n * crowd-code Extension Entry Point\n * Version 2.0 - State-based observation-action capture\n */\n\nimport * as vscode from 'vscode'\nimport * as crypto from 'crypto'\nimport { getExportPath, logToOutput, outputChannel, addToGitignore } from './utilities'\nimport {\n\tstartRecording,\n\tstopRecording,\n\tupdateStatusBarItem,\n\tpanicButton,\n\tcommands,\n\trecording,\n} from './recording'\nimport { RecordFilesProvider, type RecordFile } from './recordFilesProvider'\nimport { ActionsProvider } from './actionsProvider'\nimport {\n\tcleanupViewportCapture,\n\tcleanupTerminalCapture,\n\tcleanupFilesystemWatcher,\n} from './capture'\nimport { initializeGitProvider, cleanupGitProvider } from './gitProvider'\nimport * as fs from 'node:fs'\nimport * as path from 'node:path'\nimport { showConsentChangeDialog, ensureConsent } from './consent'\n\nexport let statusBarItem: vscode.StatusBarItem\nexport let extContext: vscode.ExtensionContext\nexport let actionsProvider: ActionsProvider\n\nfunction onConfigurationChange(event: vscode.ConfigurationChangeEvent) {\n\tif (event.affectsConfiguration('crowdCode')) {\n\t\tupdateStatusBarItem()\n\t\tgetExportPath()\n\t}\n}\n\n/**\n * Gets the full path for a file or folder\n */\nfunction getFullPath(item: RecordFile, exportPath: string): string {\n\tif (item.parentPath) {\n\t\treturn path.join(exportPath, item.parentPath, item.label)\n\t}\n\treturn path.join(exportPath, item.label)\n}\n\n/**\n * Deletes a file or folder recursively\n */\nasync function deleteFileOrFolder(filePath: string): Promise<void> {\n\ttry {\n\t\tconst stat = fs.statSync(filePath)\n\t\tif (stat.isDirectory()) {\n\t\t\tfs.rmSync(filePath, { recursive: true, force: true })\n\t\t} else {\n\t\t\tfs.unlinkSync(filePath)\n\t\t}\n\t} catch (err) {\n\t\tconsole.error('Error deleting file or folder:', err)\n\t\tthrow err\n\t}\n}\n\nexport async function activate(context: vscode.ExtensionContext): Promise<void> {\n\textContext = context\n\toutputChannel.show()\n\tlogToOutput('Activating crowd-code v2.0', 'info')\n\n\t// Generate anonymous user ID\n\tconst userName = process.env.USER || process.env.USERNAME || 'coder'\n\tconst machineId = vscode.env.machineId ?? null\n\tconst rawId = `${machineId}:${userName}`\n\tconst anonUserId = crypto.createHash('sha256').update(rawId).digest('hex')\n\n\textContext.globalState.update('userId', anonUserId)\n\n\t// Register userID display command\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand('crowd-code.showUserId', () => {\n\t\t\tconst userId = extContext.globalState.get<string>('userId')\n\t\t\tif (!userId) {\n\t\t\t\tvscode.window.showWarningMessage(\n\t\t\t\t\t'User ID not registered yet. Please wait a few seconds until the extension is fully activated.'\n\t\t\t\t)\n\t\t\t\treturn\n\t\t\t}\n\t\t\tvscode.window.showInformationMessage(`Your User ID is: ${userId}`)\n\t\t})\n\t)\n\n\t// Register Record Files Provider\n\tconst recordFilesProvider = new RecordFilesProvider()\n\tcontext.subscriptions.push(vscode.window.registerTreeDataProvider('recordFiles', recordFilesProvider))\n\n\t// Register Actions Provider\n\tactionsProvider = new ActionsProvider()\n\tcontext.subscriptions.push(vscode.window.registerTreeDataProvider('actions', actionsProvider))\n\n\t// Register refresh command\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand('crowd-code.refreshRecordFiles', () => {\n\t\t\trecordFilesProvider.refresh()\n\t\t})\n\t)\n\n\t// Register delete command\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand('crowd-code.deleteRecordFile', async (item: RecordFile) => {\n\t\t\tconst exportPath = getExportPath()\n\t\t\tif (!exportPath) {\n\t\t\t\treturn\n\t\t\t}\n\n\t\t\tconst result = await vscode.window.showWarningMessage(\n\t\t\t\t`Are you sure you want to delete ${item.label}?`,\n\t\t\t\t'Yes',\n\t\t\t\t'No'\n\t\t\t)\n\n\t\t\tif (result === 'Yes') {\n\t\t\t\ttry {\n\t\t\t\t\tconst itemPath = getFullPath(item, exportPath)\n\t\t\t\t\tawait deleteFileOrFolder(itemPath)\n\t\t\t\t\trecordFilesProvider.refresh()\n\t\t\t\t} catch (err) {\n\t\t\t\t\tvscode.window.showErrorMessage(`Error deleting ${item.label}: ${err}`)\n\t\t\t\t}\n\t\t\t}\n\t\t})\n\t)\n\n\t// Register reveal in explorer command\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand('crowd-code.revealInExplorer', (item: RecordFile) => {\n\t\t\tconst exportPath = getExportPath()\n\t\t\tif (!exportPath) {\n\t\t\t\treturn\n\t\t\t}\n\n\t\t\tconst itemPath = getFullPath(item, exportPath)\n\t\t\tvscode.commands.executeCommand('revealFileInOS', vscode.Uri.file(itemPath))\n\t\t})\n\t)\n\n\t// Register recording commands\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand(commands.startRecording, () => {\n\t\t\tstartRecording()\n\t\t})\n\t)\n\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand(commands.stopRecording, () => {\n\t\t\tstopRecording()\n\t\t})\n\t)\n\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand(commands.panicButton, () => {\n\t\t\tpanicButton()\n\t\t})\n\t)\n\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand(commands.openSettings, () => {\n\t\t\tvscode.commands.executeCommand('workbench.action.openSettings', '@ext:pdoom-org.crowd-code')\n\t\t})\n\t)\n\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand('crowd-code.addToGitignore', async () => {\n\t\t\tawait addToGitignore()\n\t\t})\n\t)\n\n\t// Register consent management command\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand('crowd-code.consent', async () => {\n\t\t\tawait showConsentChangeDialog()\n\t\t})\n\t)\n\n\t// Listen for configuration changes\n\tcontext.subscriptions.push(vscode.workspace.onDidChangeConfiguration(onConfigurationChange))\n\n\t// Initialize git provider (detects git operations for annotation)\n\tinitializeGitProvider(context)\n\n\t// Create status bar item\n\tstatusBarItem = vscode.window.createStatusBarItem(vscode.StatusBarAlignment.Right, 9000)\n\tupdateStatusBarItem()\n\tcontext.subscriptions.push(statusBarItem)\n\n\t// Ensure consent is obtained when the extension is first activated\n\tawait ensureConsent()\n\n\t// Autostart recording regardless of consent. The consent only gates data upload.\n\tlogToOutput('Autostarting recording...', 'info')\n\tstartRecording().catch((err) => logToOutput(`Autostart recording failed unexpectedly: ${err}`, 'error'))\n}\n\nexport function deactivate(): void {\n\tlogToOutput('Deactivating crowd-code v2.0', 'info')\n\n\tif (recording.isRecording) {\n\t\tstopRecording()\n\t}\n\n\t// Cleanup all modules\n\tcleanupViewportCapture()\n\tcleanupTerminalCapture()\n\tcleanupFilesystemWatcher()\n\tcleanupGitProvider()\n\n\tstatusBarItem.dispose()\n}\n",typescript,tab
+50,3604029,"src/extension.ts",3245,0,"",typescript,selection_mouse
+51,3605672,"src/extension.ts",0,0,"",typescript,selection_command
+52,3609582,"TERMINAL",0,0,"squeue",,terminal_command
+53,3609596,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  58009 yannick.ki  standard     1    32  R 2026-01-09T10:08:03 2026-01-09T10:08:03     8:41 1-00:00:00 hai001\r\n  58010 yannick.ki  standard     1    32  R 2026-01-09T10:08:03 2026-01-09T10:08:03     8:41 1-00:00:00 hai007\r\n  58011 yannick.ki  standard     1    32  R 2026-01-09T10:08:03 2026-01-09T10:08:03     8:41 1-00:00:00 hai007\r\n  58006 yannick.ki  standard     1    32  R 2026-01-09T10:06:16 2026-01-09T10:06:16    10:28 1-00:00:00 hai001\r\n  58007 yannick.ki  standard     1    32  R 2026-01-09T10:06:16 2026-01-09T10:06:16    10:28 1-00:00:00 hai001\r\n  58008 yannick.ki  standard     1    32  R 2026-01-09T10:06:16 2026-01-09T10:06:16    10:28 1-00:00:00 hai001\r\n  57998 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-09T01:09:29  9:07:15 1-00:00:00 hai008\r\n  57997 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-09T01:09:28  9:07:16 1-00:00:00 hai008\r\n  57996 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-09T00:00:43 10:16:01 1-00:00:00 hai006\r\n  57995 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-08T23:52:34 10:24:10 1-00:00:00 hai002\r\n  57993 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-08T22:30:06 11:46:38 1-00:00:00 hai006\r\n  57994 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-08T22:30:06 11:46:38 1-00:00:00 hai007\r\n  57989 yannick.ki  standard     1    32  R 2026-01-08T22:29:39 2026-01-08T22:29:40 11:47:04 1-00:00:00 hai002\r\n  57990 yannick.ki  standard     1    32  R 2026-01-08T22:29:39 2026-01-08T22:29:40 11:47:04 1-00:00:00 hai006\r\n  57883 nishant.ku  standard     3   624  R 2026-01-08T16:50:21 2026-01-08T16:50:22 17:26:22 1-00:00:00 hai[003-005]\r\n  57867 franz.sram  standard     1    16  R 2026-01-08T11:52:19 2026-01-08T11:52:19 22:24:25 1-00:00:00 hai001\r\n]0;franz.srambical@hai-login1:~/crowd-code",,terminal_output
+54,3610888,"TERMINAL",0,0,"squeue",,terminal_command
+55,3610890,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  58009 yannick.ki  standard     1    32  R 2026-01-09T10:08:03 2026-01-09T10:08:03     8:43 1-00:00:00 hai001\r\n  58010 yannick.ki  standard     1    32  R 2026-01-09T10:08:03 2026-01-09T10:08:03     8:43 1-00:00:00 hai007\r\n  58011 yannick.ki  standard     1    32  R 2026-01-09T10:08:03 2026-01-09T10:08:03     8:43 1-00:00:00 hai007\r\n  58006 yannick.ki  standard     1    32  R 2026-01-09T10:06:16 2026-01-09T10:06:16    10:30 1-00:00:00 hai001\r\n  58007 yannick.ki  standard     1    32  R 2026-01-09T10:06:16 2026-01-09T10:06:16    10:30 1-00:00:00 hai001\r\n  58008 yannick.ki  standard     1    32  R 2026-01-09T10:06:16 2026-01-09T10:06:16    10:30 1-00:00:00 hai001\r\n  57998 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-09T01:09:29  9:07:17 1-00:00:00 hai008\r\n  57997 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-09T01:09:28  9:07:18 1-00:00:00 hai008\r\n  57996 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-09T00:00:43 10:16:03 1-00:00:00 hai006\r\n  57995 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-08T23:52:34 10:24:12 1-00:00:00 hai002\r\n  57993 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-08T22:30:06 11:46:40 1-00:00:00 hai006\r\n  57994 yannick.ki  standard     1    32  R 2026-01-08T22:30:06 2026-01-08T22:30:06 11:46:40 1-00:00:00 hai007\r\n  57989 yannick.ki  standard     1    32  R 2026-01-08T22:29:39 2026-01-08T22:29:40 11:47:06 1-00:00:00 hai002\r\n  57990 yannick.ki  standard     1    32  R 2026-01-08T22:29:39 2026-01-08T22:29:40 11:47:06 1-00:00:00 hai006\r\n  57883 nishant.ku  standard     3   624  R 2026-01-08T16:50:21 2026-01-08T16:50:22 17:26:24 1-00:00:00 hai[003-005]\r\n  57867 franz.sram  standard     1    16  R 2026-01-08T11:52:19 2026-01-08T11:52:19 22:24:27 1-00:00:00 hai001\r\n]0;franz.srambical@hai-login1:~/crowd-code",,terminal_output
+56,3620374,"src/extension.ts",5773,0,"",typescript,selection_command
+57,3631771,"src/recording.ts",0,0,"",typescript,tab
+58,3632506,"src/recording.ts",13130,0,"",typescript,selection_command
+59,3633276,"src/recording.ts",0,0,"",typescript,selection_command
+60,3633400,"src/recording.ts",1581,0,"",typescript,selection_command
+61,3634908,"src/recording.ts",13147,0,"",typescript,selection_command
+62,3637182,"src/recording.ts",13193,0,"",typescript,selection_command
+63,3637426,"src/recording.ts",13219,0,"",typescript,selection_command
+64,3637454,"src/recording.ts",13250,0,"",typescript,selection_command
+65,3637488,"src/recording.ts",13252,0,"",typescript,selection_command
+66,3637526,"src/recording.ts",13302,0,"",typescript,selection_command
+67,3637562,"src/recording.ts",13350,0,"",typescript,selection_command
+68,3638166,"src/recording.ts",13352,0,"",typescript,selection_command
+69,3638360,"src/recording.ts",13378,0,"",typescript,selection_command
+70,3638828,"src/recording.ts",13352,0,"",typescript,selection_command
+71,3639150,"src/recording.ts",13355,0,"",typescript,selection_command
+72,3643092,"src/recording.ts",13350,0,"",typescript,selection_command
+73,3643333,"src/recording.ts",13305,0,"",typescript,selection_command
+74,3643365,"src/recording.ts",13255,0,"",typescript,selection_command
+75,3643499,"src/recording.ts",13250,0,"",typescript,selection_command
+76,3643643,"src/recording.ts",13222,0,"",typescript,selection_command
+77,3643791,"src/recording.ts",13196,0,"",typescript,selection_command
+78,3644014,"src/recording.ts",13200,0,"",typescript,selection_command
+79,3644575,"src/recording.ts",15914,0,"",typescript,selection_command
+80,3648307,"src/recording.ts",15964,0,"",typescript,selection_command
+81,3648437,"src/recording.ts",16002,0,"",typescript,selection_command
+82,3648798,"src/recording.ts",16031,0,"",typescript,selection_command
+83,3651543,"src/recording.ts",1828,0,"",typescript,selection_command
+84,3669878,"src/recording.ts",16031,0,"",typescript,selection_command
+85,3680364,"src/recording.ts",16028,0,"",typescript,selection_command
+86,3680604,"src/recording.ts",16028,1,"C",typescript,selection_command
+87,3680740,"src/recording.ts",16028,26,"CROWD_CODE_API_GATEWAY_URL",typescript,selection_command
+88,3681098,"src/recording.ts",16053,0,"",typescript,selection_command
+89,3687526,"TERMINAL",0,0,"echo $CROWD_CODE_API_GATEWAY_URL",,terminal_command
+90,3726577,"src/recording.ts",16122,0,"",typescript,selection_command
+91,3726691,"src/recording.ts",16128,0,"",typescript,selection_command
+92,3726818,"src/recording.ts",16130,0,"",typescript,selection_command
+93,3726976,"src/recording.ts",16168,0,"",typescript,selection_command
+94,3727367,"src/recording.ts",16167,0,"",typescript,selection_command
+95,3727547,"src/recording.ts",16154,0,"",typescript,selection_command
+96,3728294,"src/utilities.ts",0,0,"import * as vscode from 'vscode'\r\nimport * as fs from 'node:fs'\r\nimport * as path from 'node:path'\r\nimport { contributes } from '../package.json'\r\nimport { tmpdir } from 'node:os'\r\n\r\ninterface DefaultConfiguration {\r\n\t[key: string]: (typeof defaultConfiguration)[keyof typeof defaultConfiguration]\r\n}\r\n\r\nconst defaultConfiguration = contributes.configuration.properties\r\n\r\nexport const outputChannel = vscode.window.createOutputChannel('crowd-code')\r\n\r\n/**\r\n * Retrieves the configuration object for the 'crowdCode' extension.\r\n *\r\n * @returns The configuration object for the 'crowdCode' extension.\r\n */\r\nexport function getConfig() {\r\n\treturn vscode.workspace.getConfiguration('crowdCode')\r\n}\r\n\r\n/**\r\n * Creates a directory at the specified path if it does not already exist.\r\n *\r\n * @param path - The path of the directory to create.\r\n * @returns Void.\r\n */\r\nexport async function createPath(path: string) {\r\n\t// If the setting is enabled and the path is inside the workspace, add it to .gitignore\r\n\tif (\r\n\t\tgetConfig().get<boolean>('export.addToGitignore') &&\r\n\t\tgetConfig().get<string>('export.exportPath')?.startsWith('${workspaceFolder}')\r\n\t) {\r\n\t\tawait addToGitignore()\r\n\t}\r\n\r\n\tif (!fs.existsSync(path)) {\r\n\t\tfs.mkdirSync(path)\r\n\t}\r\n}\r\n\r\n/**\r\n * Retrieves the export path for the crowd-code extension, handling various scenarios such as:\r\n * - If no export path is specified, it prompts the user to reset to default or open the settings.\r\n * - If the export path starts with '${workspaceFolder}', it replaces it with the actual workspace path.\r\n * - If the export path does not exist and the 'export.createPathOutsideWorkspace' setting is false, it prompts the user to reset to default or open the settings.\r\n * - It trims, normalizes, and updates the export path in the extension settings.\r\n *\r\n * @returns The normalized and updated export path, or `undefined` if an error occurred.\r\n */\r\nexport function getExportPath(): string | undefined {\r\n\tconst exportPath = getConfig().get<string>('export.exportPath')\r\n\tlet outputExportPath = exportPath\r\n\tconst resetToDefaultMessage = 'Reset to default'\r\n\tconst openSettingsMessage = 'Open Settings'\r\n\tconst cancelMessage = 'Cancel'\r\n\r\n\t/**\r\n\t * Handles the user's selection when prompted to reset the export path to the default or open the settings.\r\n\t *\r\n\t * @param selection - The user's selection, which can be 'Reset to default', 'Open Settings', or 'Cancel'.\r\n\t * @returns Void.\r\n\t */\r\n\tfunction handleSelection(selection: string | undefined) {\r\n\t\tif (selection === resetToDefaultMessage) {\r\n\t\t\tgetConfig().update('export.exportPath', undefined, vscode.ConfigurationTarget.Global)\r\n\t\t}\r\n\t\tif (selection === 'Open Settings') {\r\n\t\t\tvscode.commands.executeCommand(\r\n\t\t\t\t'workbench.action.openSettings',\r\n\t\t\t\t'crowdCode.export.exportPath'\r\n\t\t\t)\r\n\t\t}\r\n\t}\r\n\tif (outputExportPath?.startsWith('${TMPDIR}')) {\r\n\t\tconst exportDir = tmpdir() ?? ""/tmp""\r\n\t\toutputExportPath = outputExportPath.replace('${TMPDIR}', exportDir)\r\n\t\ttry {\r\n\t\t\tif (!fs.existsSync(outputExportPath)) {\r\n\t\t\t\tfs.mkdirSync(outputExportPath, { recursive: true })\r\n\t\t\t}\r\n\t\t} catch (err) {\r\n\t\t\tconst errorMessage = vscode.l10n.t('Failed to create export path: {0}', String(err));\r\n\t\t\tvscode.window.showErrorMessage(errorMessage);\r\n\t\t\tlogToOutput(errorMessage, 'error');\r\n\t\t\treturn\r\n\t\t}\r\n\t}\r\n\r\n\tif (!outputExportPath) {\r\n\t\tconst exportPathNotFoundMessage = 'No export path specified'\r\n\t\tvscode.window\r\n\t\t\t.showErrorMessage(\r\n\t\t\t\texportPathNotFoundMessage,\r\n\t\t\t\tresetToDefaultMessage,\r\n\t\t\t\topenSettingsMessage,\r\n\t\t\t\tcancelMessage\r\n\t\t\t)\r\n\t\t\t.then(selection => handleSelection(selection))\r\n\t\tlogToOutput(exportPathNotFoundMessage, 'error')\r\n\t\treturn\r\n\t}\r\n\r\n\tif (outputExportPath?.startsWith('${workspaceFolder}')) {\r\n\t\tconst workspacePath = vscode.workspace.workspaceFolders?.[0].uri.fsPath\r\n\t\tif (!workspacePath) {\r\n\t\t\tconst errorMessage = 'No workspace folder found'\r\n\t\t\tvscode.window.showErrorMessage(errorMessage)\r\n\t\t\tlogToOutput(errorMessage, 'error')\r\n\t\t\treturn\r\n\t\t}\r\n\t\toutputExportPath = outputExportPath.replace('${workspaceFolder}', workspacePath)\r\n\t\tcreatePath(outputExportPath)\r\n\t} else {\r\n\t\tif (\r\n\t\t\t!fs.existsSync(outputExportPath) &&\r\n\t\t\tgetConfig().get<boolean>('export.createPathOutsideWorkspace', false) === false\r\n\t\t) {\r\n\t\t\tconst exportPathNotFoundMessage = 'Export path does not exist'\r\n\t\t\tvscode.window\r\n\t\t\t\t.showErrorMessage(\r\n\t\t\t\t\texportPathNotFoundMessage,\r\n\t\t\t\t\tresetToDefaultMessage,\r\n\t\t\t\t\topenSettingsMessage,\r\n\t\t\t\t\tcancelMessage\r\n\t\t\t\t)\r\n\t\t\t\t// deepcode ignore PromiseNotCaughtGeneral: catch method not available\r\n\t\t\t\t.then(selection => handleSelection(selection))\r\n\t\t\tlogToOutput(exportPathNotFoundMessage, 'error')\r\n\t\t\treturn\r\n\t\t}\r\n\t\tcreatePath(outputExportPath)\r\n\t}\r\n\r\n\toutputExportPath = outputExportPath.trim()\r\n\toutputExportPath = outputExportPath.replaceAll('\\', '/')\r\n\tif (!outputExportPath.endsWith('/')) {\r\n\t\toutputExportPath += '/'\r\n\t}\r\n\r\n\tif (path.sep === '/') {\r\n\t\toutputExportPath = outputExportPath.replaceAll('/', path.sep)\r\n\t}\r\n\treturn outputExportPath\r\n}\r\n\r\nexport function setDefaultOptions() {\r\n\tconst config = getConfig()\r\n\tfor (const [key, value] of Object.entries(defaultConfiguration)) {\r\n\t\tconst configKey = key.replace('crowdCode.', '')\r\n\t\tif ('default' in value) {\r\n\t\t\tconfig.update(configKey, value.default, vscode.ConfigurationTarget.Workspace)\r\n\t\t}\r\n\t}\r\n}\r\n\r\n/**\r\n * Logs a message to the output channel with a timestamp and type.\r\n *\r\n * @param {string} message - The message to be logged.\r\n * @param {'info' | 'success' | 'error'} [type='info'] - The type of the log message.\r\n */\r\nexport function logToOutput(message: string, type: 'info' | 'success' | 'error' = 'info') {\r\n\tconst time = new Date().toLocaleTimeString()\r\n\r\n\toutputChannel.appendLine(`${time} [${type}] ${message}`)\r\n\tconsole.log(message)\r\n}\r\n\r\n/**\r\n * Generates a file name based on the current date and time.\r\n * @param date - The date to use for generating the file name.\r\n * @param isExport - Whether the file is being exported.\r\n * @param customName - Optional custom name for the folder.\r\n * @returns The generated file name.\r\n */\r\nexport function generateBaseFilePath(\r\n\tdate: Date | null,\r\n\tisExport = false,\r\n\tcustomName?: string, \r\n\tsessionId?: string\r\n): string | undefined {\r\n\tif (!date) {\r\n\t\treturn\r\n\t}\r\n\tconst year = date.getFullYear()\r\n\tconst month = (date.getMonth() + 1).toString().padStart(2, '0')\r\n\tconst day = date.getDate().toString().padStart(2, '0')\r\n\tconst hours = date.getHours().toString().padStart(2, '0')\r\n\tconst minutes = date.getMinutes().toString().padStart(2, '0')\r\n\tconst seconds = date.getSeconds().toString().padStart(2, '0')\r\n\tconst milliseconds = date.getMilliseconds().toString().padStart(2, '0')\r\n\r\n\tconst timestamp = `${year}_${month}_${day}-${hours}.${minutes}.${seconds}.${milliseconds}`\r\n\tconst default_name = sessionId ? `crowd-code-${sessionId}-${timestamp}` : `crowd-code-${timestamp}`\r\n\tconst folderName = customName ? `${customName}-${timestamp}` : default_name\r\n\tconst fileName = isExport ? 'recording' : 'source'\r\n\r\n\treturn `${folderName}/${fileName}`\r\n}\r\n\r\n/**\r\n * Retrieves the language identifier of the currently active text editor.\r\n *\r\n * @return {string} The language identifier of the active text editor\r\n */\r\nexport function getEditorLanguage(): string {\r\n\tconst editor = vscode.window.activeTextEditor\r\n\tif (editor) {\r\n\t\tconsole.log(editor.document.languageId)\r\n\t\treturn editor.document.languageId\r\n\t}\r\n\treturn ''\r\n}\r\n\r\n/**\r\n * Gets the relative path of the active text editor's file.\r\n * @returns A string representing the relative path of the active text editor's file.\r\n */\r\nexport function getEditorFileName(): string {\r\n\treturn vscode.workspace.asRelativePath(vscode.window.activeTextEditor?.document.fileName ?? '')\r\n}\r\n\r\n/**\r\n * Displays a notification with progress in VS Code.\r\n * @param title - The title of the notification.\r\n */\r\nexport function notificationWithProgress(title: string): void {\r\n\tvscode.window.withProgress(\r\n\t\t{\r\n\t\t\tlocation: vscode.ProgressLocation.Notification,\r\n\t\t\ttitle: title,\r\n\t\t\tcancellable: false,\r\n\t\t},\r\n\t\tprogress => {\r\n\t\t\treturn new Promise<void>(resolve => {\r\n\t\t\t\tconst times = 1.5 * 1000\r\n\t\t\t\tconst timeout = 50\r\n\t\t\t\tconst increment = (100 / times) * timeout\r\n\t\t\t\tfor (let i = 0; i <= times; i++) {\r\n\t\t\t\t\tsetTimeout(() => {\r\n\t\t\t\t\t\tprogress.report({ increment: increment })\r\n\t\t\t\t\t\tif (i === times / timeout) {\r\n\t\t\t\t\t\t\tresolve()\r\n\t\t\t\t\t\t}\r\n\t\t\t\t\t}, timeout * i)\r\n\t\t\t\t}\r\n\t\t\t})\r\n\t\t}\r\n\t)\r\n}\r\n\r\n/**\r\n * Formats a time value in seconds to a display string.\r\n * @param seconds - The number of seconds.\r\n * @returns A string representing the formatted time.\r\n */\r\nexport function formatDisplayTime(seconds: number): string {\r\n\tconst hours = Math.floor(seconds / 3600)\r\n\tconst minutes = Math.floor((seconds % 3600) / 60)\r\n\tconst remainingSeconds = seconds % 60\r\n\r\n\tlet timeString = ''\r\n\r\n\tif (hours > 0) {\r\n\t\ttimeString += `${hours.toString().padStart(2, '0')}:`\r\n\t}\r\n\r\n\ttimeString += `${minutes.toString().padStart(2, '0')}:${remainingSeconds\r\n\t\t.toString()\r\n\t\t.padStart(2, '0')}`\r\n\r\n\treturn timeString\r\n}\r\n\r\n/**\r\n * Formats a time value in milliseconds to an SRT time string.\r\n * @param milliseconds - The number of milliseconds.\r\n * @returns A string representing the formatted SRT time.\r\n */\r\nexport function formatSrtTime(milliseconds: number): string {\r\n\tconst seconds = Math.floor(milliseconds / 1000)\r\n\tconst hours = Math.floor(seconds / 3600)\r\n\tconst minutes = Math.floor((seconds % 3600) / 60)\r\n\tconst remainingSeconds = seconds % 60\r\n\tconst remainingMilliseconds = milliseconds % 1000\r\n\r\n\treturn `${hours.toString().padStart(2, '0')}:${minutes\r\n\t\t.toString()\r\n\t\t.padStart(2, '0')}:${remainingSeconds.toString().padStart(2, '0')},${remainingMilliseconds\r\n\t\t.toString()\r\n\t\t.padStart(3, '0')}`\r\n}\r\n\r\n/**\r\n * Escapes special characters in a string for CSV compatibility.\r\n * @param editorText - The text to escape.\r\n * @returns A string with escaped characters.\r\n */\r\nexport function escapeString(editorText: string | undefined): string {\r\n\tif (editorText === undefined) {\r\n\t\treturn ''\r\n\t}\r\n\treturn editorText\r\n\t\t.replace(/""/g, '""""')\r\n\t\t.replace(/\r\n/g, '\\r\\n')\r\n\t\t.replace(/\n/g, '\\n')\r\n\t\t.replace(/\r/g, '\\r')\r\n\t\t.replace(/\t/g, '\\t')\r\n}\r\n\r\n/**\r\n * Removes double quotes at the start and end of a text string.\r\n * @param text - The text to process.\r\n * @returns A string without surrounding double quotes.\r\n */\r\nexport function removeDoubleQuotes(text: string): string {\r\n\treturn text.replace(/^""(.*)""$/, '$1')\r\n}\r\n\r\n/**\r\n * Unescape special characters in a string.\r\n * @param text - The text to unescape.\r\n * @returns A string with unescaped characters.\r\n */\r\nexport function unescapeString(text: string): string {\r\n\treturn text\r\n\t\t.replace(/""""/g, '""')\r\n\t\t.replace(/\\r\\n/g, '\r\n')\r\n\t\t.replace(/\\n/g, '\n')\r\n\t\t.replace(/\\r/g, '\r')\r\n\t\t.replace(/\\t/g, '\t')\r\n}\r\n\r\n/**\r\n * Adds the export path to .gitignore if it doesn't exist.\r\n * @returns true if the path was added, false if it already exists or if there was an error\r\n */\r\nexport async function addToGitignore(): Promise<boolean> {\r\n\tconst workspaceFolder = vscode.workspace.workspaceFolders?.[0]\r\n\tif (!workspaceFolder) {\r\n\t\tvscode.window.showErrorMessage('No workspace found')\r\n\t\treturn false\r\n\t}\r\n\r\n\tconst gitignorePath = path.join(workspaceFolder.uri.fsPath, '.gitignore')\r\n\tconst exportPath = getConfig().get<string>('export.exportPath')\r\n\r\n\tif (!exportPath) {\r\n\t\tvscode.window.showErrorMessage('No export path specified')\r\n\t\treturn false\r\n\t}\r\n\r\n\t// Get the relative path from workspace folder\r\n\tlet relativePath = exportPath\r\n\tif (exportPath.startsWith('${workspaceFolder}')) {\r\n\t\trelativePath = exportPath.replace('${workspaceFolder}', '').replace(/\\/g, '/')\r\n\t}\r\n\t// Remove leading and trailing slashes\r\n\trelativePath = relativePath.replace(/^\/+|\/+$/g, '')\r\n\r\n\ttry {\r\n\t\tlet content = ''\r\n\t\tif (fs.existsSync(gitignorePath)) {\r\n\t\t\tcontent = fs.readFileSync(gitignorePath, 'utf8')\r\n\t\t\t// Check if the path is already in .gitignore\r\n\t\t\tif (content.split('\n').some(line => line.trim() === relativePath)) {\r\n\t\t\t\tvscode.window.showInformationMessage('Export path already in .gitignore')\r\n\t\t\t\treturn false\r\n\t\t\t}\r\n\t\t\t// Add a newline if the file doesn't end with one\r\n\t\t\tif (!content.endsWith('\n')) {\r\n\t\t\t\tcontent += '\n'\r\n\t\t\t}\r\n\t\t}\r\n\t\tcontent = `${content}${relativePath}\n`\r\n\t\tfs.writeFileSync(gitignorePath, content)\r\n\t\tvscode.window.showInformationMessage('Export path added to .gitignore')\r\n\t\treturn true\r\n\t} catch (err) {\r\n\t\tconsole.error('Error updating .gitignore:', err)\r\n\t\tvscode.window.showErrorMessage('Error updating .gitignore')\r\n\t\treturn false\r\n\t}\r\n}\r\n",typescript,tab
+97,3728294,"src/utilities.ts",1915,0,"",typescript,selection_command
+98,3770790,"src/utilities.ts",0,0,"",typescript,tab
+99,3774865,"src/utilities.ts",1970,0,"",typescript,selection_command
+100,3776588,"src/utilities.ts",2036,0,"",typescript,selection_command
+101,3776681,"src/utilities.ts",2072,0,"",typescript,selection_command
+102,3785530,"src/utilities.ts",2123,0,"",typescript,selection_command
+103,3785780,"src/utilities.ts",2169,0,"",typescript,selection_command
+104,3785806,"src/utilities.ts",2186,0,"",typescript,selection_command
+105,3785860,"src/utilities.ts",2191,0,"",typescript,selection_command
+106,3785874,"src/utilities.ts",2210,0,"",typescript,selection_command
+107,3785923,"src/utilities.ts",2306,0,"",typescript,selection_command
+108,3785951,"src/utilities.ts",2325,0,"",typescript,selection_command
+109,3786002,"src/utilities.ts",2434,0,"",typescript,selection_command
+110,3786007,"src/utilities.ts",2441,0,"",typescript,selection_command
+111,3786055,"src/utilities.ts",2460,0,"",typescript,selection_command
+112,3786094,"src/utilities.ts",2520,0,"",typescript,selection_command
+113,3786123,"src/utilities.ts",2566,0,"",typescript,selection_command
+114,3786161,"src/utilities.ts",2642,0,"",typescript,selection_command
+115,3786191,"src/utilities.ts",2661,0,"",typescript,selection_command
+116,3786230,"src/utilities.ts",2701,0,"",typescript,selection_command
+117,3786328,"src/utilities.ts",2737,0,"",typescript,selection_command
+118,3786373,"src/utilities.ts",2775,0,"",typescript,selection_command
+119,3786437,"src/utilities.ts",2797,0,"",typescript,selection_command
+120,3786437,"src/utilities.ts",2802,0,"",typescript,selection_command
+121,3786566,"src/utilities.ts",2806,0,"",typescript,selection_command
+122,3786739,"src/utilities.ts",2825,0,"",typescript,selection_command
+123,3786796,"src/utilities.ts",2876,0,"",typescript,selection_command
+124,3787135,"src/utilities.ts",2825,0,"",typescript,selection_command
+125,3792919,"src/utilities.ts",2902,0,"",typescript,selection_command
+126,3793203,"src/utilities.ts",2921,0,"",typescript,selection_command
+127,3793365,"src/utilities.ts",3002,0,"",typescript,selection_command
+128,3793520,"src/utilities.ts",3041,0,"",typescript,selection_command
+129,3793856,"src/utilities.ts",3314,0,"",typescript,selection_command
+130,3795411,"src/utilities.ts",3341,0,"",typescript,selection_command
+131,3795422,"src/utilities.ts",3405,0,"",typescript,selection_command
+132,3795964,"src/utilities.ts",3422,0,"",typescript,selection_command
+133,3796099,"src/utilities.ts",3445,0,"",typescript,selection_command
+134,3796231,"src/utilities.ts",3477,0,"",typescript,selection_command
+135,3796427,"src/utilities.ts",3505,0,"",typescript,selection_command
+136,3796656,"src/utilities.ts",3531,0,"",typescript,selection_command
+137,3796847,"src/utilities.ts",3547,0,"",typescript,selection_command
+138,3797005,"src/utilities.ts",3556,0,"",typescript,selection_command
+139,3797434,"src/utilities.ts",3607,0,"",typescript,selection_command
+140,3798571,"src/utilities.ts",3556,0,"",typescript,selection_command
+141,3798813,"src/utilities.ts",3547,0,"",typescript,selection_command
+142,3798837,"src/utilities.ts",3531,0,"",typescript,selection_command
+143,3798882,"src/utilities.ts",3505,0,"",typescript,selection_command
+144,3798910,"src/utilities.ts",3477,0,"",typescript,selection_command
+145,3798937,"src/utilities.ts",3445,0,"",typescript,selection_command
+146,3798979,"src/utilities.ts",3422,0,"",typescript,selection_command
+147,3799020,"src/utilities.ts",3405,0,"",typescript,selection_command
+148,3799299,"src/utilities.ts",3341,0,"",typescript,selection_command
+149,3806349,"src/utilities.ts",3314,0,"",typescript,selection_command
+150,3806725,"src/utilities.ts",2025,0,"",typescript,selection_command
+151,3807898,"src/utilities.ts",2042,0,"",typescript,selection_command
+152,3808045,"src/utilities.ts",2044,0,"",typescript,selection_command
+153,3809724,"src/utilities.ts",1978,0,"",typescript,selection_command
+154,3810139,"src/utilities.ts",1974,0,"",typescript,selection_command
+155,3810306,"src/utilities.ts",1972,0,"",typescript,selection_command
+156,3810465,"src/utilities.ts",1961,0,"",typescript,selection_command
+157,3810612,"src/utilities.ts",1955,0,"",typescript,selection_command
+158,3810909,"src/utilities.ts",1961,0,"",typescript,selection_command
+159,3811146,"src/utilities.ts",1972,0,"",typescript,selection_command
+160,3811189,"src/utilities.ts",1974,0,"",typescript,selection_command
+161,3811212,"src/utilities.ts",1983,0,"",typescript,selection_command
+162,3811250,"src/utilities.ts",1986,0,"",typescript,selection_command
+163,3811389,"src/utilities.ts",1989,0,"",typescript,selection_command
+164,3811641,"src/utilities.ts",1990,0,"",typescript,selection_command
+165,3814065,"package.json",0,0,"{\n  ""name"": ""crowd-code"",\n  ""displayName"": ""crowd-code"",\n  ""description"": ""Record your code editing activity in VS Code"",\n  ""version"": ""1.1.3"",\n  ""publisher"": ""pdoom-org"",\n  ""icon"": ""icon.png"",\n  ""engines"": {\n    ""vscode"": ""^1.89.0""\n  },\n  ""main"": ""./out/extension.js"",\n  ""categories"": [\n    ""Other"",\n    ""Education""\n  ],\n  ""activationEvents"": [\n    ""onStartupFinished""\n  ],\n  ""contributes"": {\n    ""commands"": [\n      {\n        ""command"": ""crowd-code.startRecording"",\n        ""title"": ""Start Recording"",\n        ""icon"": ""$(play)""\n      },\n      {\n        ""command"": ""crowd-code.stopRecording"",\n        ""title"": ""Stop Recording"",\n        ""icon"": ""$(stop)""\n      },\n      {\n        ""command"": ""crowd-code.panicButton"",\n        ""title"": ""Panic Button"",\n        ""icon"": ""$(refresh)""\n      },\n      {\n        ""command"": ""crowd-code.openSettings"",\n        ""title"": ""Open Settings"",\n        ""icon"": ""$(settings)""\n      },\n      {\n        ""command"": ""crowd-code.refreshRecordFiles"",\n        ""title"": ""Refresh"",\n        ""icon"": ""$(refresh)""\n      },\n      {\n        ""command"": ""crowd-code.deleteRecordFile"",\n        ""title"": ""Delete"",\n        ""icon"": ""$(trash)""\n      },\n      {\n        ""command"": ""crowd-code.revealInExplorer"",\n        ""title"": ""Reveal in File Explorer"",\n        ""icon"": ""$(folder-opened)""\n      },\n      {\n        ""command"": ""crowd-code.addToGitignore"",\n        ""title"": ""Add to .gitignore"",\n        ""icon"": ""$(git-ignore)""\n      },\n      {\n        ""command"": ""crowd-code.showUserId"",\n        ""title"": ""Show User ID""\n      },\n      {\n        ""command"": ""crowd-code.consent"",\n        ""title"": ""crowd-code: Manage Data Collection Consent""\n      }\n    ],\n    ""viewsContainers"": {\n      ""activitybar"": [\n        {\n          ""id"": ""crowd-code"",\n          ""title"": ""crowd-code"",\n          ""icon"": ""icon.svg""\n        }\n      ]\n    },\n    ""views"": {\n      ""crowd-code"": [\n        {\n          ""id"": ""actions"",\n          ""name"": ""Actions"",\n          ""icon"": ""icon.svg"",\n          ""contextualTitle"": ""Actions""\n        },\n        {\n          ""id"": ""recordFiles"",\n          ""name"": ""Record Files"",\n          ""icon"": ""icon.svg"",\n          ""contextualTitle"": ""Record Files""\n        }\n      ]\n    },\n    ""menus"": {\n      ""view/title"": [\n        {\n          ""command"": ""crowd-code.refreshRecordFiles"",\n          ""when"": ""view == recordFiles"",\n          ""group"": ""navigation""\n        },\n        {\n          ""command"": ""crowd-code.addToGitignore"",\n          ""when"": ""view == actions"",\n          ""group"": ""navigation""\n        },\n        {\n          ""command"": ""crowd-code.consent"",\n          ""when"": ""view == actions"",\n          ""group"": ""navigation""\n        }\n      ],\n      ""view/item/context"": [\n        {\n          ""command"": ""crowd-code.deleteRecordFile"",\n          ""when"": ""view == recordFiles"",\n          ""group"": ""inline""\n        },\n        {\n          ""command"": ""crowd-code.revealInExplorer"",\n          ""when"": ""view == recordFiles"",\n          ""group"": ""inline""\n        },\n        {\n          ""command"": ""crowd-code.deleteRecordFile"",\n          ""when"": ""view == recordFiles"",\n          ""group"": ""1_modification""\n        },\n        {\n          ""command"": ""crowd-code.revealInExplorer"",\n          ""when"": ""view == recordFiles"",\n          ""group"": ""2_workspace""\n        }\n      ]\n    },\n    ""configuration"": {\n      ""title"": ""crowd-code Configuration"",\n      ""properties"": {\n        ""crowdCode.export.exportPath"": {\n          ""type"": ""string"",\n          ""default"": ""${TMPDIR}/crowd-code/"",\n          ""markdownDescription"": ""Path where recordings will be exported. Use '${workspaceFolder}' to refer to the workspace folder."",\n          ""order"": 0\n        },\n        ""crowdCode.export.createPathOutsideWorkspace"": {\n          ""type"": ""boolean"",\n          ""default"": true,\n          ""description"": ""Whether to create the export path if it doesn't exist, even if it's outside the workspace folder."",\n          ""order"": 1\n        },\n        ""crowdCode.export.addToGitignore"": {\n          ""type"": ""boolean"",\n          ""default"": false,\n          ""description"": ""Whether to add the export path to .gitignore when creating it."",\n          ""order"": 2\n        },\n        ""crowdCode.export.exportFormats"": {\n          ""type"": ""array"",\n          ""items"": {\n            ""type"": ""string"",\n            ""enum"": [\n              ""JSON"",\n              ""SRT""\n            ]\n          },\n          ""uniqueItems"": true,\n          ""default"": [\n            ""JSON"",\n            ""SRT""\n          ],\n          ""description"": ""Formats to export the recording in."",\n          ""order"": 3\n        },\n        ""crowdCode.recording.askFolderName"": {\n          ""type"": ""boolean"",\n          ""default"": false,\n          ""description"": ""Ask for a custom folder name before starting a recording"",\n          ""order"": 4\n        },\n        ""crowdCode.appearance.minimalMode"": {\n          ""type"": ""boolean"",\n          ""default"": false,\n          ""description"": ""Whether to show only the icon in the status bar when recording is stopped."",\n          ""order"": 5\n        },\n        ""crowdCode.appearance.showTimer"": {\n          ""type"": ""boolean"",\n          ""default"": true,\n          ""description"": ""Whether to show the timer in the status bar when recording."",\n          ""order"": 6\n        }\n      }\n    }\n  },\n  ""repository"": {\n    ""type"": ""git"",\n    ""url"": ""git://github.com/p-doom/crowd-code.git""\n  },\n  ""scripts"": {\n    ""vscode:prepublish"": ""npm run package"",\n    ""ovsx:publish"": ""ovsx publish"",\n    ""vsce:package"": ""vsce package"",\n    ""release"": ""npm run package && vsce package"",\n    ""compile"": ""webpack"",\n    ""watch"": ""webpack --watch"",\n    ""package"": ""webpack --mode production --devtool hidden-source-map"",\n    ""compile-tests"": ""tsc -p . --outDir out"",\n    ""watch-tests"": ""tsc -p . -w --outDir out"",\n    ""pretest"": ""npm run compile-tests && npm run compile && npm run lint"",\n    ""lint"": ""eslint src --ext ts"",\n    ""test"": ""vscode-test""\n  },\n  ""devDependencies"": {\n    ""@types/diff"": ""^5.2.0"",\n    ""@types/mocha"": ""^10.0.6"",\n    ""@types/node"": ""18.x"",\n    ""@types/vscode"": ""^1.89.0"",\n    ""@typescript-eslint/eslint-plugin"": ""^7.11.0"",\n    ""@typescript-eslint/parser"": ""^7.11.0"",\n    ""@vscode/l10n-dev"": ""^0.0.35"",\n    ""@vscode/test-cli"": ""^0.0.9"",\n    ""@vscode/test-electron"": ""^2.4.0"",\n    ""@vscode/vsce"": ""^2.24.0"",\n    ""eslint"": ""^8.57.0"",\n    ""ts-loader"": ""^9.5.1"",\n    ""typescript"": ""^5.4.5"",\n    ""webpack"": ""^5.91.0"",\n    ""webpack-cli"": ""^5.1.4""\n  },\n  ""dependencies"": {\n    ""axios"": ""^1.7.2"",\n    ""diff"": ""^5.2.0"",\n    ""ignore"": ""^5.3.0"",\n    ""lru-cache"": ""^10.0.0""\n  }\n}\n",json,tab
+166,3816150,"package.json",3398,0,"",json,selection_command
+167,3817439,"package.json",3439,0,"",json,selection_command
+168,3817558,"package.json",3467,0,"",json,selection_command
+169,3824565,"src/utilities.ts",0,0,"",typescript,tab
+170,3829281,"TERMINAL",0,0,"module load nodejs",,terminal_command
+171,3829330,"TERMINAL",0,0,"]633;C",,terminal_output
+172,3829918,"TERMINAL",0,0,"]0;franz.srambical@hai-login1:~/crowd-code",,terminal_output
+173,3832208,"TERMINAL",0,0,"npm run compile",,terminal_command
+174,3832260,"TERMINAL",0,0,"]633;C",,terminal_output
+175,3834306,"TERMINAL",0,0,"\r\n> crowd-code@1.1.3 compile\r\n> webpack\r\n\r\n",,terminal_output
+176,3836512,"TERMINAL",0,0,"    [webpack-cli] Compiler starting... \r\n    [webpack-cli] Compiler is using config: '/fast/home/franz.srambical/crowd-code/webpack.config.js'\r\n",,terminal_output
+177,3839557,"TERMINAL",0,0,"    [webpack-cli] Compiler finished\r\nasset extension.js 763 KiB [emitted] (name: main) 1 related asset\r\nruntime modules 670 bytes 3 modules\r\nmodules by path ./node_modules/ 655 KiB 84 modules\r\nmodules by path ./src/ 91 KiB\r\n  modules by path ./src/*.ts 69.2 KiB\r\n    ./src/extension.ts 8.13 KiB [built] [code generated]\r\n    ./src/utilities.ts 14.6 KiB [built] [code generated]\r\n    ./src/recording.ts 23.1 KiB [built] [code generated]\r\n    ./src/consent.ts 5.32 KiB [built] [code generated]\r\n    + 3 modules\r\n  modules by path ./src/capture/*.ts 21.7 KiB\r\n    ./src/capture/index.ts 2.55 KiB [built] [code generated]\r\n    ./src/capture/viewportCapture.ts 4.59 KiB [built] [code generated]\r\n    ./src/capture/terminalCapture.ts 7.83 KiB [built] [code generated]\r\n    ./src/capture/filesystemWatcher.ts 6.76 KiB [built] [code generated]\r\n+ 19 modules\r\nwebpack 5.99.9 compiled successfully in 3042 ms\r\nnpm notice \r\nnpm notice New major version of npm available! 10.5.2 -> 11.7.0\r\nnpm notice Changelog: https://github.com/npm/cli/releases/tag/v11.7.0\r\nnpm notice Run npm install -g npm@11.7.0 to update!\r\nnpm notice \r\n]0;franz.srambical@hai-login1:~/crowd-code",,terminal_output
+178,3854475,"TERMINAL",0,0,"vsce package",,terminal_command
+179,3854524,"TERMINAL",0,0,"]633;C",,terminal_output
+180,3855602,"TERMINAL",0,0,"^C\r\n]0;franz.srambical@hai-login1:~/crowd-code",,terminal_output
+181,3858608,"package.json",0,0,"",json,tab
+182,3859114,"package.json",0,0,"",json,selection_command
+183,3860544,"package.json",2,0,"",json,selection_command
+184,3860790,"package.json",26,0,"",json,selection_command
+185,3860821,"package.json",57,0,"",json,selection_command
+186,3860858,"package.json",122,0,"",json,selection_command
+187,3861316,"package.json",144,0,"",json,selection_command
+188,3861584,"package.json",122,0,"",json,selection_command
+189,3862134,"package.json",124,0,"",json,selection_command
+190,3862296,"package.json",125,0,"",json,selection_command
+191,3862765,"package.json",132,0,"",json,selection_command
+192,3862765,"package.json",135,0,"",json,selection_command
+193,3862845,"package.json",136,0,"",json,selection_command
+194,3863728,"package.json",136,1,"2",json,content
+195,3864130,"package.json",137,0,"",json,selection_command
+196,3864694,"package.json",138,0,"",json,selection_command
+197,3865192,"package.json",138,1,"0",json,content
+198,3865388,"package.json",139,0,"",json,selection_command
+199,3865549,"package.json",140,0,"",json,selection_command
+200,3866182,"package.json",140,1,"0",json,content
+201,3871113,"TERMINAL",0,0,"vsce package",,terminal_command
+202,3871159,"TERMINAL",0,0,"]633;C",,terminal_output
+203,3873150,"TERMINAL",0,0,"Executing prepublish script 'npm run vscode:prepublish'...\r\n",,terminal_output
+204,3874101,"package.json",661,0,"",json,selection_command
+205,3874250,"package.json",1305,0,"",json,selection_command
+206,3874380,"TERMINAL",0,0,"\r\n> crowd-code@2.0.0 vscode:prepublish\r\n> npm run package\r\n\r\n",,terminal_output
+207,3874705,"package.json",1858,0,"",json,selection_command
+208,3874857,"package.json",2486,0,"",json,selection_command
+209,3875267,"package.json",3143,0,"",json,selection_command
+210,3875427,"package.json",4128,0,"",json,selection_command
+211,3875630,"TERMINAL",0,0,"\r\n> crowd-code@2.0.0 package\r\n> webpack --mode production --devtool hidden-source-map\r\n\r\n",,terminal_output
+212,3875785,"package.json",4793,0,"",json,selection_command
+213,3875979,"package.json",5566,0,"",json,selection_command
+214,3876292,"package.json",6417,0,"",json,selection_command
+215,3877711,"TERMINAL",0,0,"    [webpack-cli] Compiler starting... \r\n    [webpack-cli] Compiler is using config: '/fast/home/franz.srambical/crowd-code/webpack.config.js'\r\n",,terminal_output
+216,3881309,"TERMINAL",0,0,"    [webpack-cli] Compiler finished\r\nasset extension.js 321 KiB [emitted] [minimized] (name: main) 2 related assets\r\nruntime modules 670 bytes 3 modules\r\norphan modules 84 bytes [orphan] 2 modules\r\nmodules by path ./node_modules/ 655 KiB 84 modules\r\nmodules by path ./src/ 91 KiB\r\n  modules by path ./src/*.ts 69.2 KiB\r\n    ./src/extension.ts 8.13 KiB [built] [code generated]\r\n    ./src/utilities.ts 14.6 KiB [built] [code generated]\r\n    ./src/recording.ts 23.1 KiB [built] [code generated]\r\n    + 4 modules\r\n  modules by path ./src/capture/*.ts 21.7 KiB\r\n    ./src/capture/index.ts 2.55 KiB [built] [code generated]\r\n    ./src/capture/viewportCapture.ts 4.59 KiB [built] [code generated]\r\n    ./src/capture/terminalCapture.ts 7.83 KiB [built] [code generated]\r\n    ./src/capture/filesystemWatcher.ts 6.76 KiB [built] [code generated]\r\n+ 17 modules\r\nwebpack 5.99.9 compiled successfully in 3544 ms\r\nnpm notice \r\nnpm notice New major version of npm available! 10.5.2 -> 11.7.0\r\nnpm notice Changelog: https://github.com/npm/cli/releases/tag/v11.7.0\r\nnpm notice Run npm install -g npm@11.7.0 to update!\r\nnpm notice \r\nnpm notice \r\nnpm notice New major version of npm available! 10.5.2 -> 11.7.0\r\nnpm notice Changelog: https://github.com/npm/cli/releases/tag/v11.7.0\r\nnpm notice Run npm install -g npm@11.7.0 to update!\r\nnpm notice \r\n",,terminal_output
+217,3883586,"TERMINAL",0,0," INFO  Files included in the VSIX:\r\ncrowd-code-2.0.0.vsix\r\n├─ [Content_Types].xml \r\n├─ extension.vsixmanifest \r\n└─ extension/\r\n   ├─ LICENSE.md [1.05 KB]\r\n   ├─ biome.json [0.11 KB]\r\n   ├─ changelog.md \r\n   ├─ icon.png [513.5 KB]\r\n   ├─ icon.svg [206.57 KB]\r\n   ├─ package.json [6.38 KB]\r\n   ├─ readme.md \r\n   ├─ .github/\r\n   │  └─ workflows/\r\n   │     └─ release.yml [2.02 KB]\r\n   ├─ img/\r\n   │  └─ preview.gif [338.02 KB]\r\n   ├─ media/\r\n   │  ├─ main.js [0.19 KB]\r\n   │  └─ style.css [1 KB]\r\n   ├─ out/\r\n   │  ├─ extension.js [321.21 KB]\r\n   │  └─ extension.js.LICENSE.txt [0.31 KB]\r\n   └─ test-workspace/\r\n      └─ .vscode/\r\n         └─ settings.json [0.26 KB]\r\n\r\n",,terminal_output
+218,3884498,"TERMINAL",0,0," DONE  Packaged: /fast/home/franz.srambical/crowd-code/crowd-code-2.0.0.vsix (16 files, 963.21 KB)\r\n]0;franz.srambical@hai-login1:~/crowd-code",,terminal_output
+219,3885615,"package.json",0,0,"",json,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-f48e876f-ca91-4d87-816a-ddb62ba190ce1767782830349-2026_01_07-11.47.24.290/source.csv

@@ -0,0 +1,204 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"/home/franz.srambical/crowd-code/src/extension.ts",0,0,"import * as vscode from 'vscode'\nimport * as crypto from 'crypto'\nimport { getExportPath, logToOutput, outputChannel, addToGitignore } from './utilities'\nimport {\n\tupdateStatusBarItem,\n\tstartRecording,\n\tstopRecording,\n\tisCurrentFileExported,\n\tcommands,\n\trecording,\n\taddToFileQueue,\n\tbuildCsvRow,\n\tappendToFile,\n\tpanicButton,\n} from './recording'\nimport { ChangeType, CSVRowBuilder } from './types'\nimport { RecordFilesProvider, type RecordFile } from './recordFilesProvider'\nimport { ActionsProvider } from './actionsProvider'\nimport { initializeGitProvider, cleanupGitProvider } from './gitProvider'\nimport * as fs from 'node:fs'\nimport * as path from 'node:path'\nimport { showConsentChangeDialog, ensureConsent, hasConsent } from './consent'\n\nexport let statusBarItem: vscode.StatusBarItem\nexport let extContext: vscode.ExtensionContext\nexport let actionsProvider: ActionsProvider\n\nfunction onConfigurationChange(event: vscode.ConfigurationChangeEvent) {\n\tif (event.affectsConfiguration('crowdCode')) {\n\t\tupdateStatusBarItem()\n\t\tgetExportPath()\n\t}\n}\n\n/**\n * Gets the full path for a file or folder\n * @param item - The tree item representing the file or folder\n * @param exportPath - The base export path\n * @returns The full path to the file or folder\n */\nfunction getFullPath(item: RecordFile, exportPath: string): string {\n\t// If the item has a parent path (file inside a folder), construct the full path\n\tif (item.parentPath) {\n\t\treturn path.join(exportPath, item.parentPath, item.label)\n\t}\n\t// Otherwise, it's a root item\n\treturn path.join(exportPath, item.label)\n}\n\n/**\n * Deletes a file or folder recursively\n * @param filePath - The path to the file or folder to delete\n */\nasync function deleteFileOrFolder(filePath: string): Promise<void> {\n\ttry {\n\t\tconst stat = fs.statSync(filePath)\n\t\tif (stat.isDirectory()) {\n\t\t\t// Delete directory and its contents recursively\n\t\t\tfs.rmSync(filePath, { recursive: true, force: true })\n\t\t} else {\n\t\t\t// Delete single file\n\t\t\tfs.unlinkSync(filePath)\n\t\t}\n\t} catch (err) {\n\t\tconsole.error('Error deleting file or folder:', err)\n\t\tthrow err\n\t}\n}\n\nexport async function activate(context: vscode.ExtensionContext): Promise<void> {\n\textContext = context\n\toutputChannel.show()\n\tlogToOutput('Activating crowd-code', 'info')\n\n\t// Save anonUserId globally for user to copy\n\tconst userName = process.env.USER || process.env.USERNAME || ""coder"";\n\tconst machineId = vscode.env.machineId ?? null;\n\tconst rawId = `${machineId}:${userName}`;\n\tconst anonUserId = crypto.createHash('sha256').update(rawId).digest('hex') as string;\n\n\textContext.globalState.update('userId', anonUserId);\n\n\t// Register userID display\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand('crowd-code.showUserId', () => {\n\t\t\tconst userId = extContext.globalState.get<string>('userId');\n\t\t\tif (!userId) {\n\t\t\t\tvscode.window.showWarningMessage(""User ID not registered yet. Please wait a few seconds until the extension is fully activated."");\n\t\t\t\treturn;\n\t\t\t}\n\t\t\tvscode.window.showInformationMessage(`Your User ID is: ${userId}`);\n\t\t}))\n\n\n\t// Register Record Files Provider\n\tconst recordFilesProvider = new RecordFilesProvider()\n\tcontext.subscriptions.push(\n\t\tvscode.window.registerTreeDataProvider('recordFiles', recordFilesProvider)\n\t)\n\n\t// Register Actions Provider\n\tactionsProvider = new ActionsProvider()\n\tcontext.subscriptions.push(vscode.window.registerTreeDataProvider('actions', actionsProvider))\n\n\t// Register refresh command\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand('crowd-code.refreshRecordFiles', () => {\n\t\t\trecordFilesProvider.refresh()\n\t\t})\n\t)\n\n\t// Register delete command\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand(\n\t\t\t'crowd-code.deleteRecordFile',\n\t\t\tasync (item: RecordFile) => {\n\t\t\t\tconst exportPath = getExportPath()\n\t\t\t\tif (!exportPath) {\n\t\t\t\t\treturn\n\t\t\t\t}\n\n\t\t\t\tconst result = await vscode.window.showWarningMessage(\n\t\t\t\t\t`Are you sure you want to delete ${item.label}?`,\n\t\t\t\t\t'Yes',\n\t\t\t\t\t'No'\n\t\t\t\t)\n\n\t\t\t\tif (result === 'Yes') {\n\t\t\t\t\ttry {\n\t\t\t\t\t\tconst itemPath = getFullPath(item, exportPath)\n\t\t\t\t\t\tawait deleteFileOrFolder(itemPath)\n\t\t\t\t\t\trecordFilesProvider.refresh()\n\t\t\t\t\t} catch (err) {\n\t\t\t\t\t\tvscode.window.showErrorMessage(`Error deleting ${item.label}: ${err}`)\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\t\t)\n\t)\n\n\t// Register reveal in explorer command\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand('crowd-code.revealInExplorer', (item: RecordFile) => {\n\t\t\tconst exportPath = getExportPath()\n\t\t\tif (!exportPath) {\n\t\t\t\treturn\n\t\t\t}\n\n\t\t\tconst itemPath = getFullPath(item, exportPath)\n\t\t\tvscode.commands.executeCommand('revealFileInOS', vscode.Uri.file(itemPath))\n\t\t})\n\t)\n\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand(commands.startRecording, () => {\n\t\t\tstartRecording()\n\t\t})\n\t)\n\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand(commands.stopRecording, () => {\n\t\t\tstopRecording()\n\t\t})\n\t)\n\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand(commands.panicButton, () => {\n\t\t\tpanicButton()\n\t\t})\n\t)\n\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand(commands.openSettings, () => {\n\t\t\tvscode.commands.executeCommand(\n\t\t\t\t'workbench.action.openSettings',\n\t\t\t\t'@ext:MattiaConsiglio.crowd-code'\n\t\t\t)\n\t\t})\n\t)\n\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand('crowd-code.addToGitignore', async () => {\n\t\t\tawait addToGitignore()\n\t\t})\n\t)\n\n\t// Register consent management command\n\tcontext.subscriptions.push(\n\t\tvscode.commands.registerCommand('crowd-code.consent', async () => {\n\t\t\tawait showConsentChangeDialog()\n\t\t})\n\t)\n\n\n\tcontext.subscriptions.push(vscode.workspace.onDidChangeConfiguration(onConfigurationChange))\n\n\tvscode.window.onDidChangeActiveTextEditor(editor => {\n\t\tupdateStatusBarItem()\n\t\tif (editor && recording.isRecording) {\n\t\t\tif (isCurrentFileExported()) {\n\t\t\t\treturn\n\t\t\t}\n\t\t\tconst currentFileUri = editor.document.uri.toString()\n\t\t\tlet tabEventText = ''\n\n\t\t\tif (recording.activatedFiles) {\n\t\t\t\tif (!recording.activatedFiles.has(currentFileUri)) {\n\t\t\t\t\ttabEventText = editor.document.getText()\n\t\t\t\t\trecording.activatedFiles.add(currentFileUri)\n\t\t\t\t}\n\t\t\t} else {\n\t\t\t\tthrow new Error(""Warning: recording.activatedFiles was not available during TAB event logging."")\n\t\t\t}\n\n\t\t\trecording.sequence++\n\t\t\taddToFileQueue(\n\t\t\t\tbuildCsvRow({\n\t\t\t\t\tsequence: recording.sequence,\n\t\t\t\t\trangeOffset: 0,\n\t\t\t\t\trangeLength: 0,\n\t\t\t\t\ttext: tabEventText,\n\t\t\t\t\ttype: ChangeType.TAB,\n\t\t\t\t})\n\t\t\t)\n\t\t\tappendToFile()\n\t\t\tactionsProvider.setCurrentFile(editor.document.fileName)\n\t\t}\n\t})\n\n\tcontext.subscriptions.push(\n\t\tvscode.window.onDidChangeTextEditorSelection(event => {\n\t\t\tif (recording.isRecording && event.textEditor === vscode.window.activeTextEditor) {\n\t\t\t\tif (isCurrentFileExported()) {\n\t\t\t\t\treturn\n\t\t\t\t}\n\n\t\t\t\tconst editor = event.textEditor\n\t\t\t\t// For simplicity, we'll log the primary selection.\n\t\t\t\tconst selection = event.selections[0]\n\t\t\t\tif (!selection) {\n\t\t\t\t\treturn\n\t\t\t\t}\n\n\t\t\t\tconst selectedText = editor.document.getText(selection)\n\t\t\t\tlet changeType: string\n\n\t\t\t\tswitch (event.kind) {\n\t\t\t\t\tcase vscode.TextEditorSelectionChangeKind.Keyboard:\n\t\t\t\t\t\tchangeType = ChangeType.SELECTION_KEYBOARD\n\t\t\t\t\t\tbreak\n\t\t\t\t\tcase vscode.TextEditorSelectionChangeKind.Mouse:\n\t\t\t\t\t\tchangeType = ChangeType.SELECTION_MOUSE\n\t\t\t\t\t\tbreak\n\t\t\t\t\tcase vscode.TextEditorSelectionChangeKind.Command:\n\t\t\t\t\t\tchangeType = ChangeType.SELECTION_COMMAND\n\t\t\t\t\t\tbreak\n\t\t\t\t\tdefault:\n\t\t\t\t\t\tthrow new TypeError(""Unknown selection change kind."")\n\t\t\t\t}\n\n\t\t\t\trecording.sequence++\n\t\t\t\tconst csvRowParams: CSVRowBuilder = {\n\t\t\t\t\tsequence: recording.sequence,\n\t\t\t\t\trangeOffset: editor.document.offsetAt(selection.start),\n\t\t\t\t\trangeLength: editor.document.offsetAt(selection.end) - editor.document.offsetAt(selection.start),\n\t\t\t\t\ttext: selectedText,\n\t\t\t\t\ttype: changeType,\n\t\t\t\t}\n\t\t\t\taddToFileQueue(buildCsvRow(csvRowParams))\n\t\t\t\tappendToFile()\n\t\t\t\tactionsProvider.setCurrentFile(editor.document.fileName)\n\t\t\t}\n\t\t})\n\t)\n\n\tcontext.subscriptions.push(\n\t\tvscode.window.onDidChangeActiveTerminal((terminal: vscode.Terminal | undefined) => {\n\t\t\tif (terminal && recording.isRecording) {\n\t\t\t\tif (isCurrentFileExported()) {\n\t\t\t\t\treturn\n\t\t\t\t}\n\t\t\t\trecording.sequence++\n\t\t\t\taddToFileQueue(\n\t\t\t\t\tbuildCsvRow({\n\t\t\t\t\t\tsequence: recording.sequence,\n\t\t\t\t\t\trangeOffset: 0,\n\t\t\t\t\t\trangeLength: 0,\n\t\t\t\t\t\ttext: terminal.name,\n\t\t\t\t\t\ttype: ChangeType.TERMINAL_FOCUS,\n\t\t\t\t\t})\n\t\t\t\t)\n\t\t\t\tappendToFile()\n\t\t\t\tactionsProvider.setCurrentFile(`Terminal: ${terminal.name}`)\n\t\t\t}\n\t\t})\n\t)\n\n\tcontext.subscriptions.push(\n\t\tvscode.window.onDidStartTerminalShellExecution(async (event: vscode.TerminalShellExecutionStartEvent) => {\n\t\t\tif (recording.isRecording) {\n\t\t\t\tif (isCurrentFileExported()) {\n\t\t\t\t\treturn\n\t\t\t\t}\n\t\t\t\tconst commandLine = event.execution.commandLine.value\n\t\t\t\trecording.sequence++\n\t\t\t\taddToFileQueue(\n\t\t\t\t\tbuildCsvRow({\n\t\t\t\t\t\tsequence: recording.sequence,\n\t\t\t\t\t\trangeOffset: 0,\n\t\t\t\t\t\trangeLength: 0,\n\t\t\t\t\t\ttext: commandLine,\n\t\t\t\t\t\ttype: ChangeType.TERMINAL_COMMAND,\n\t\t\t\t\t})\n\t\t\t\t)\n\t\t\t\tappendToFile()\n\n\t\t\t\tconst stream = event.execution.read()\n\t\t\t\tfor await (const data of stream) {\n\t\t\t\t\trecording.sequence++\n\t\t\t\t\taddToFileQueue(\n\t\t\t\t\t\tbuildCsvRow({ sequence: recording.sequence, rangeOffset: 0, rangeLength: 0, text: data, type: ChangeType.TERMINAL_OUTPUT })\n\t\t\t\t\t)\n\t\t\t\t\tappendToFile()\n\t\t\t\t}\n\t\t\t}\n\t\t})\n\t)\n\n\tstatusBarItem = vscode.window.createStatusBarItem(vscode.StatusBarAlignment.Right, 9000)\n\tupdateStatusBarItem()\n\tcontext.subscriptions.push(statusBarItem)\n\n\t// Ensure consent is obtained when the extension is first activated\n\tawait ensureConsent()\n\n\t// Autostart recording regardless of consent. The consent only gates data upload.\n\tstartRecording().catch(err => logToOutput(`Autostart recording failed unexpectedly: ${err}`, 'error'))\n\n\t// Initialize git provider for branch checkout detection\n\tinitializeGitProvider()\n}\n\nexport function deactivate(): void {\n\tlogToOutput('Deactivating crowd-code', 'info')\n\tif (recording.isRecording) {\n\t\tstopRecording()\n\t}\n\tcleanupGitProvider()\n\tstatusBarItem.dispose()\n}\n",typescript,tab
+2,245,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"11:47:24 AM [info] Activating crowd-code\n11:47:24 AM [info] Recording started\n11:47:24 AM [info] Initializing git provider using file system watchers...\n11:47:24 AM [info] Git repository found\n11:47:24 AM [info] Git provider initialized successfully\n11:47:24 AM [info] Initial git state: [object Object]\n",Log,tab
+3,1192819,"/home/franz.srambical/crowd-code/src/extension.ts",0,0,"",typescript,tab
+4,1970993,"Untitled-1",0,0,"",plaintext,tab
+5,1985966,"Untitled-1",0,0,"t",plaintext,content
+6,1985968,"Untitled-1",1,0,"",plaintext,selection_keyboard
+7,1986002,"Untitled-1",1,0,"h",plaintext,content
+8,1986003,"Untitled-1",2,0,"",plaintext,selection_keyboard
+9,1986306,"Untitled-1",2,0,"i",plaintext,content
+10,1986307,"Untitled-1",3,0,"",plaintext,selection_keyboard
+11,1986308,"Untitled-1",3,0,"s",plaintext,content
+12,1986308,"Untitled-1",4,0,"",plaintext,selection_keyboard
+13,1986310,"Untitled-1",4,0," ",plaintext,content
+14,1986310,"Untitled-1",5,0,"",plaintext,selection_keyboard
+15,1986335,"Untitled-1",5,0,"i",plaintext,content
+16,1986335,"Untitled-1",6,0,"",plaintext,selection_keyboard
+17,1986495,"Untitled-1",6,0,"s",plaintext,content
+18,1986495,"Untitled-1",7,0,"",plaintext,selection_keyboard
+19,1986551,"Untitled-1",7,0," ",plaintext,content
+20,1986551,"Untitled-1",8,0,"",plaintext,selection_keyboard
+21,1986702,"Untitled-1",8,0,"a",plaintext,content
+22,1986702,"Untitled-1",9,0,"",plaintext,selection_keyboard
+23,1986795,"Untitled-1",9,0," ",plaintext,content
+24,1986795,"Untitled-1",10,0,"",plaintext,selection_keyboard
+25,1986993,"Untitled-1",10,0,"t",plaintext,content
+26,1986993,"Untitled-1",11,0,"",plaintext,selection_keyboard
+27,1987062,"Untitled-1",11,0,"e",plaintext,content
+28,1987063,"Untitled-1",12,0,"",plaintext,selection_keyboard
+29,1987192,"Untitled-1",12,0,"s",plaintext,content
+30,1987193,"Untitled-1",13,0,"",plaintext,selection_keyboard
+31,1987205,"Untitled-1",13,0,"t",plaintext,content
+32,1987205,"Untitled-1",14,0,"",plaintext,selection_keyboard
+33,1989800,"Untitled-1",13,0,"",plaintext,selection_command
+34,1992387,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",0,0,"--- TRACE STARTED ---\n--- TRACE STOPPED ---\n",Log,tab
+35,1993422,"Untitled-1",0,0,"",plaintext,tab
+36,1993814,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",0,0,"",Log,tab
+37,1996907,"Untitled-1",0,0,"",plaintext,tab
+38,1999056,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",0,0,"",Log,tab
+39,2008936,"Untitled-1",0,0,"",plaintext,tab
+40,2010694,"Untitled-1",0,14,"",plaintext,content
+41,2010700,"/home/franz.srambical/crowd-code/src/extension.ts",0,0,"",typescript,tab
+42,2012683,"/home/franz.srambical/crowd-code/src/extension.ts",9983,0,"",typescript,selection_command
+43,2016622,"/home/franz.srambical/crowd-code/src/extension.ts",9983,0,"\n",typescript,content
+44,2017013,"/home/franz.srambical/crowd-code/src/extension.ts",9984,0,"t",typescript,content
+45,2017014,"/home/franz.srambical/crowd-code/src/extension.ts",9985,0,"",typescript,selection_keyboard
+46,2017087,"/home/franz.srambical/crowd-code/src/extension.ts",9985,0,"e",typescript,content
+47,2017088,"/home/franz.srambical/crowd-code/src/extension.ts",9986,0,"",typescript,selection_keyboard
+48,2017236,"/home/franz.srambical/crowd-code/src/extension.ts",9986,0,"s",typescript,content
+49,2017236,"/home/franz.srambical/crowd-code/src/extension.ts",9987,0,"",typescript,selection_keyboard
+50,2017244,"/home/franz.srambical/crowd-code/src/extension.ts",9987,0,"t",typescript,content
+51,2017245,"/home/franz.srambical/crowd-code/src/extension.ts",9988,0,"",typescript,selection_keyboard
+52,2019059,"/home/franz.srambical/crowd-code/src/extension.ts",9987,0,"",typescript,selection_command
+53,2022542,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",0,0,"",Log,tab
+54,2023365,"/home/franz.srambical/crowd-code/src/extension.ts",0,0,"",typescript,tab
+55,2029549,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",0,0,"",Log,tab
+56,2033491,"/home/franz.srambical/crowd-code/src/extension.ts",0,0,"",typescript,tab
+57,2034092,"/home/franz.srambical/crowd-code/src/extension.ts",9983,5,"",typescript,content
+58,2036358,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",0,0,"",Log,tab
+59,2039013,"/home/franz.srambical/crowd-code/src/extension.ts",0,0,"",typescript,tab
+60,2061351,"/home/franz.srambical/test_extension/test/src/extension.ts",0,0,"import * as vscode from 'vscode'\n\nlet output: vscode.OutputChannel\nlet tracingEnabled = false\n\nfunction isInWorkspace(uri: vscode.Uri): boolean {\n    const folders = vscode.workspace.workspaceFolders\n    if (!folders) return false\n    return folders.some(f => uri.fsPath.startsWith(f.uri.fsPath))\n}\n\nfunction shouldIgnore(uri: vscode.Uri): boolean {\n    const p = uri.fsPath\n\n    // Ignore VS Code internals and extension noise\n    if (p.includes(`${pathSep}.vscode${pathSep}`)) return true\n    if (p.includes(`${pathSep}node_modules${pathSep}`)) return true\n    if (p.includes(`${pathSep}.git${pathSep}`)) return true\n\n    return false\n}\n\nconst pathSep = require('path').sep\n\nexport function activate(context: vscode.ExtensionContext) {\n    output = vscode.window.createOutputChannel('Edit Trace Probe')\n\n    const log = (msg: string) => {\n        if (!tracingEnabled) return\n        output.appendLine(`[${new Date().toISOString()}] ${msg}`)\n    }\n\n    // ---- Commands ----\n    context.subscriptions.push(\n        vscode.commands.registerCommand('editTraceProbe.start', () => {\n            tracingEnabled = true\n            output.show(true)\n            output.appendLine('--- TRACE STARTED ---')\n        })\n    )\n\n    context.subscriptions.push(\n        vscode.commands.registerCommand('editTraceProbe.stop', () => {\n            output.appendLine('--- TRACE STOPPED ---')\n            tracingEnabled = false\n        })\n    )\n\n    context.subscriptions.push(\n        vscode.commands.registerCommand('editTraceProbe.showOutput', () => {\n            output.show(true)\n        })\n    )\n\n    // ---- Text document changes ----\n    context.subscriptions.push(\n        vscode.workspace.onDidChangeTextDocument(event => {\n            const uri = event.document.uri\n            if (!isInWorkspace(uri)) return\n            if (shouldIgnore(uri)) return\n\n            log(\n                `TEXT_CHANGE\n  uri=${uri.toString()}\n  version=${event.document.version}\n  changes=${event.contentChanges.length}\n  active=${vscode.window.activeTextEditor?.document === event.document}`\n            )\n        })\n    )\n\n    // ---- Document open ----\n    context.subscriptions.push(\n        vscode.workspace.onDidOpenTextDocument(doc => {\n            const uri = doc.uri\n            if (!isInWorkspace(uri)) return\n            if (shouldIgnore(uri)) return\n\n            log(\n                `DOC_OPEN\n  uri=${uri.toString()}\n  version=${doc.version}\n  language=${doc.languageId}`\n            )\n        })\n    )\n\n    // ---- Filesystem watcher (workspace only) ----\n    const watcher = vscode.workspace.createFileSystemWatcher('**/*')\n    context.subscriptions.push(watcher)\n\n    watcher.onDidChange(uri => {\n        if (!isInWorkspace(uri)) return\n        if (shouldIgnore(uri)) return\n        log(`FS_CHANGE uri=${uri.toString()}`)\n    })\n\n    watcher.onDidCreate(uri => {\n        if (!isInWorkspace(uri)) return\n        if (shouldIgnore(uri)) return\n        log(`FS_CREATE uri=${uri.toString()}`)\n    })\n\n    watcher.onDidDelete(uri => {\n        if (!isInWorkspace(uri)) return\n        if (shouldIgnore(uri)) return\n        log(`FS_DELETE uri=${uri.toString()}`)\n    })\n}\n",typescript,tab
+61,2068022,"/home/franz.srambical/test_extension/test/src/extension.ts",376,0,"",typescript,selection_command
+62,2092367,"/home/franz.srambical/test_extension/test/src/extension.ts",1916,0,"",typescript,selection_command
+63,2106840,"/home/franz.srambical/test_extension/test/src/extension.ts",2819,0,"",typescript,selection_command
+64,2115138,"/home/franz.srambical/test_extension/test/src/extension.ts",0,0,"",typescript,selection_command
+65,2115991,"/home/franz.srambical/test_extension/test/src/extension.ts",3152,0,"",typescript,selection_command
+66,2116681,"/home/franz.srambical/test_extension/test/src/extension.ts",3150,0,"",typescript,selection_command
+67,2116920,"/home/franz.srambical/test_extension/test/src/extension.ts",3143,0,"",typescript,selection_command
+68,2116952,"/home/franz.srambical/test_extension/test/src/extension.ts",3096,0,"",typescript,selection_command
+69,2116986,"/home/franz.srambical/test_extension/test/src/extension.ts",3058,0,"",typescript,selection_command
+70,2117019,"/home/franz.srambical/test_extension/test/src/extension.ts",3018,0,"",typescript,selection_command
+71,2117053,"/home/franz.srambical/test_extension/test/src/extension.ts",2985,0,"",typescript,selection_command
+72,2117090,"/home/franz.srambical/test_extension/test/src/extension.ts",2984,0,"",typescript,selection_command
+73,2117121,"/home/franz.srambical/test_extension/test/src/extension.ts",2977,0,"",typescript,selection_command
+74,2117149,"/home/franz.srambical/test_extension/test/src/extension.ts",2930,0,"",typescript,selection_command
+75,2117186,"/home/franz.srambical/test_extension/test/src/extension.ts",2892,0,"",typescript,selection_command
+76,2117220,"/home/franz.srambical/test_extension/test/src/extension.ts",2852,0,"",typescript,selection_command
+77,2117250,"/home/franz.srambical/test_extension/test/src/extension.ts",2819,0,"",typescript,selection_command
+78,2117289,"/home/franz.srambical/test_extension/test/src/extension.ts",2818,0,"",typescript,selection_command
+79,2117318,"/home/franz.srambical/test_extension/test/src/extension.ts",2811,0,"",typescript,selection_command
+80,2117353,"/home/franz.srambical/test_extension/test/src/extension.ts",2764,0,"",typescript,selection_command
+81,2117388,"/home/franz.srambical/test_extension/test/src/extension.ts",2726,0,"",typescript,selection_command
+82,2117420,"/home/franz.srambical/test_extension/test/src/extension.ts",2686,0,"",typescript,selection_command
+83,2117454,"/home/franz.srambical/test_extension/test/src/extension.ts",2653,0,"",typescript,selection_command
+84,2117484,"/home/franz.srambical/test_extension/test/src/extension.ts",2652,0,"",typescript,selection_command
+85,2117521,"/home/franz.srambical/test_extension/test/src/extension.ts",2612,0,"",typescript,selection_command
+86,2117553,"/home/franz.srambical/test_extension/test/src/extension.ts",2543,0,"",typescript,selection_command
+87,2214387,"/home/franz.srambical/crowd-code/src/extension.ts",0,0,"",typescript,tab
+88,2228653,"crates/cli/src/main.rs",0,0,"//! CLI tool for serializing crowd-pilot IDE interaction data.\n//!\n//! This tool processes CSV session files and outputs JSONL format suitable for\n//! Miles SFT training. It uses the HuggingFace tokenizers Rust library for\n//! accurate token counting.\n\nuse std::path::PathBuf;\n\nuse clap::{Parser, Subcommand};\nuse tokenizers::Tokenizer as HfTokenizer;\n\nuse crowd_pilot_serializer_core::{\n    default_system_prompt,\n    git_pull_analysis::{analyze_all_sessions, AnalysisResult},\n    pipeline::{PipelineConfig, PipelineResult},\n    process_all_sessions, write_jsonl_output, Tokenizer,\n};\n\n/// CLI tool for crowd-pilot IDE interaction data.\n#[derive(Parser, Debug)]\n#[command(name = ""crowd-pilot-serialize"")]\n#[command(author, version, about, long_about = None)]\nstruct Cli {\n    #[command(subcommand)]\n    command: Commands,\n}\n\n#[derive(Subcommand, Debug)]\nenum Commands {\n    /// Serialize CSV sessions to Miles JSONL format\n    Serialize(SerializeArgs),\n    /// Analyze sessions for git pull data corruption issues\n    AnalyzeGitPull(AnalyzeGitPullArgs),\n}\n\n/// Arguments for the serialize subcommand.\n#[derive(Parser, Debug)]\nstruct SerializeArgs {\n    /// Root directory containing CSV session files\n    #[arg(long)]\n    csv_root: PathBuf,\n\n    /// Output directory for JSONL files\n    #[arg(long)]\n    output_dir: PathBuf,\n\n    /// HuggingFace tokenizer model name or path\n    #[arg(long)]\n    tokenizer: String,\n\n    /// Maximum tokens per conversation chunk\n    #[arg(long, default_value = ""8192"")]\n    max_tokens_per_conversation: usize,\n\n    /// Maximum tokens per message\n    #[arg(long, default_value = ""2048"")]\n    max_tokens_per_message: usize,\n\n    /// Minimum messages required to keep a conversation\n    #[arg(long, default_value = ""5"")]\n    min_conversation_messages: usize,\n\n    /// Viewport radius (lines above/below cursor)\n    #[arg(long, default_value = ""10"")]\n    viewport_radius: usize,\n\n    /// Coalesce radius for grouping nearby edits\n    #[arg(long, default_value = ""5"")]\n    coalesce_radius: usize,\n\n    /// Fraction of sessions for validation (0.0-1.0)\n    #[arg(long, default_value = ""0.1"")]\n    val_ratio: f64,\n\n    /// Custom system prompt (optional)\n    #[arg(long)]\n    system_prompt: Option<String>,\n}\n\n/// Arguments for the analyze-git-pull subcommand.\n#[derive(Parser, Debug)]\nstruct AnalyzeGitPullArgs {\n    /// Root directory containing CSV session files\n    #[arg(long)]\n    csv_root: PathBuf,\n\n    /// Output JSON file for detailed results (optional)\n    #[arg(long)]\n    output: Option<PathBuf>,\n\n    /// Show detailed per-session information\n    #[arg(long, default_value = ""false"")]\n    verbose: bool,\n}\n\n/// Wrapper around HuggingFace tokenizers for token counting and truncation.\n///\n/// This uses the Rust-native tokenizers library, which is `Send + Sync`\n/// and enables true parallel tokenization without the Python GIL.\nstruct RustTokenizer {\n    inner: HfTokenizer,\n}\n\nimpl RustTokenizer {\n    /// Load a HuggingFace tokenizer from a model name or path.\n    fn load(model_name: &str) -> Result<Self, Box<dyn std::error::Error>> {\n        let inner = HfTokenizer::from_pretrained(model_name, None)\n            .map_err(|e| e as Box<dyn std::error::Error>)?;\n        Ok(Self { inner })\n    }\n}\n\nimpl Tokenizer for RustTokenizer {\n    fn count_tokens(&self, text: &str) -> usize {\n        self.inner\n            .encode(text, false)\n            .expect(""Failed to encode text with tokenizer"")\n            .get_ids()\n            .len()\n    }\n\n    fn truncate_to_max_tokens(&self, text: &str, max_tokens: usize) -> String {\n        let encoding = self.inner\n            .encode(text, false)\n            .expect(""Failed to encode text with tokenizer"");\n        \n        let ids = encoding.get_ids();\n        if ids.len() <= max_tokens {\n            return text.to_string();\n        }\n        \n        let truncated_ids: Vec<u32> = ids[..max_tokens].to_vec();\n        self.inner\n            .decode(&truncated_ids, true)\n            .expect(""Failed to decode truncated tokens"")\n    }\n}\n\nfn main() -> Result<(), Box<dyn std::error::Error>> {\n    let cli = Cli::parse();\n\n    match cli.command {\n        Commands::Serialize(args) => run_serialize(args),\n        Commands::AnalyzeGitPull(args) => run_analyze_git_pull(args),\n    }\n}\n\nfn run_serialize(args: SerializeArgs) -> Result<(), Box<dyn std::error::Error>> {\n    println!(""Loading tokenizer from {}..."", args.tokenizer);\n    let tokenizer = RustTokenizer::load(&args.tokenizer)?;\n\n    let config = PipelineConfig {\n        max_tokens_per_conversation: args.max_tokens_per_conversation,\n        max_tokens_per_message: args.max_tokens_per_message,\n        min_conversation_messages: args.min_conversation_messages,\n        viewport_radius: args.viewport_radius,\n        coalesce_radius: args.coalesce_radius,\n        val_ratio: args.val_ratio,\n    };\n\n    println!(""Processing CSV files from {:?}..."", args.csv_root);\n    let session_results = process_all_sessions(\n        &args.csv_root,\n        &tokenizer,\n        &config,\n    )?;\n\n    let total_sessions = session_results.len();\n    println!(""Processed {} sessions"", total_sessions);\n\n    let default_prompt = default_system_prompt(args.viewport_radius);\n    let system_prompt = args.system_prompt.as_deref().unwrap_or(&default_prompt);\n\n    println!(""Writing output to {:?}..."", args.output_dir);\n    let result: PipelineResult = write_jsonl_output(\n        session_results,\n        &args.output_dir,\n        args.val_ratio,\n        system_prompt,\n    )?;\n\n    let metadata_path = args.output_dir.join(""metadata.json"");\n    let metadata = serde_json::json!({\n        ""config"": {\n            ""csv_root"": args.csv_root.to_string_lossy(),\n            ""output_dir"": args.output_dir.to_string_lossy(),\n            ""tokenizer"": args.tokenizer,\n            ""max_tokens_per_conversation"": args.max_tokens_per_conversation,\n            ""max_tokens_per_message"": args.max_tokens_per_message,\n            ""min_conversation_messages"": args.min_conversation_messages,\n            ""viewport_radius"": args.viewport_radius,\n            ""coalesce_radius"": args.coalesce_radius,\n            ""val_ratio"": args.val_ratio,\n        },\n        ""counts"": {\n            ""total_sessions"": result.total_sessions,\n            ""total_conversations"": result.total_conversations,\n            ""train_conversations"": result.train_conversations,\n            ""val_conversations"": result.val_conversations,\n        },\n        ""stats"": {\n            ""total_messages"": result.total_messages,\n            ""total_tokens"": result.total_tokens,\n            ""avg_messages_per_conversation"": if result.total_conversations > 0 {\n                result.total_messages as f64 / result.total_conversations as f64\n            } else {\n                0.0\n            },\n            ""avg_tokens_per_conversation"": if result.total_conversations > 0 {\n                result.total_tokens as f64 / result.total_conversations as f64\n            } else {\n                0.0\n            },\n        },\n        ""files"": {\n            ""train_path"": args.output_dir.join(""training.jsonl"").to_string_lossy(),\n            ""val_path"": args.output_dir.join(""validation.jsonl"").to_string_lossy(),\n        },\n    });\n    std::fs::write(&metadata_path, serde_json::to_string_pretty(&metadata)?)?;\n\n    println!(""\n[summary]"");\n    println!(""  Total sessions processed: {}"", result.total_sessions);\n    println!(""  Train conversations: {}"", result.train_conversations);\n    println!(""  Val conversations: {}"", result.val_conversations);\n    println!(""  Total messages: {}"", result.total_messages);\n    println!(""  Total tokens: {}"", result.total_tokens);\n    println!(""  Output: {:?}/{{training,validation}}.jsonl"", args.output_dir);\n    println!(""  Metadata: {:?}"", metadata_path);\n\n    Ok(())\n}\n\nfn run_analyze_git_pull(args: AnalyzeGitPullArgs) -> Result<(), Box<dyn std::error::Error>> {\n    println!(""Analyzing sessions for git pull corruption..."");\n    println!(""CSV root: {:?}"", args.csv_root);\n    println!();\n\n    let result: AnalysisResult = analyze_all_sessions(&args.csv_root)?;\n\n    // Print summary\n    println!(""\n╔══════════════════════════════════════════════════════════════╗"");\n    println!(""║           GIT PULL CORRUPTION ANALYSIS RESULTS              ║"");\n    println!(""╠══════════════════════════════════════════════════════════════╣"");\n    println!(""║ Total sessions analyzed:        {:>8}                    ║"", result.total_sessions);\n    println!(""║ Sessions with terminal git pull:{:>8}                    ║"", result.sessions_with_git_pull);\n    println!(""║ Corrupted sessions (detected):  {:>8}                    ║"", result.corrupted_sessions);\n    println!(""╠══════════════════════════════════════════════════════════════╣"");\n    println!(""║ Total git pull commands:        {:>8}                    ║"", result.total_git_pulls);\n    println!(""║ Git pulls with parseable output:{:>8}                    ║"", result.parseable_git_pulls);\n    println!(""╠══════════════════════════════════════════════════════════════╣"");\n    println!(""║ Total affected files:           {:>8}                    ║"", result.total_affected_files);\n    println!(""║ Total affected edit events:     {:>8}                    ║"", result.total_affected_edits);\n    println!(""╠══════════════════════════════════════════════════════════════╣"");\n    println!(""║ CORRUPTION RATE (lower bound):  {:>7.2}%                    ║"", result.corruption_rate * 100.0);\n    println!(""╚══════════════════════════════════════════════════════════════╝"");\n\n    println!(""\n[note] This analysis only detects terminal-based git pulls."");\n    println!(""       VS Code UI git pulls (~33% of users) are undetectable."");\n    println!(""       Actual corruption rate may be ~1.5x higher."");\n\n    // Optionally write detailed results\n    if let Some(output_path) = &args.output {\n        let json = serde_json::to_string_pretty(&result)?;\n        std::fs::write(output_path, json)?;\n        println!(""\nDetailed results written to: {:?}"", output_path);\n    }\n\n    // Optionally print corrupted session paths\n    if args.verbose && !result.corrupted_session_paths.is_empty() {\n        println!(""\nCorrupted sessions:"");\n        for path in &result.corrupted_session_paths {\n            println!(""  - {}"", path);\n        }\n    }\n\n    Ok(())\n}\n",rust,tab
+89,2229296,"crates/cli/src/main.rs",10299,0,"",rust,selection_command
+90,2236040,"crates/cli/src/main.rs",10299,0,"t",rust,content
+91,2236041,"crates/cli/src/main.rs",10300,0,"",rust,selection_keyboard
+92,2236095,"crates/cli/src/main.rs",10300,0,"e",rust,content
+93,2236096,"crates/cli/src/main.rs",10301,0,"",rust,selection_keyboard
+94,2236245,"crates/cli/src/main.rs",10301,0,"s",rust,content
+95,2236245,"crates/cli/src/main.rs",10302,0,"",rust,selection_keyboard
+96,2236255,"crates/cli/src/main.rs",10302,0,"t",rust,content
+97,2236256,"crates/cli/src/main.rs",10303,0,"",rust,selection_keyboard
+98,2239709,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",0,0,"",Log,tab
+99,2245117,"crates/cli/src/main.rs",0,0,"",rust,tab
+100,2245969,"crates/cli/src/main.rs",10302,0,"",rust,selection_command
+101,2246389,"crates/cli/src/main.rs",10299,4,"",rust,content
+102,2248665,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",0,0,"",Log,tab
+103,2252574,"crates/cli/src/main.rs",0,0,"",rust,tab
+104,2254933,"/home/franz.srambical/test_extension/test/src/extension.ts",0,0,"",typescript,tab
+105,2274242,"/home/franz.srambical/crowd-code/src/extension.ts",0,0,"",typescript,tab
+106,2284300,"/home/franz.srambical/test_extension/test/src/extension.ts",0,0,"",typescript,tab
+107,2351869,"/home/franz.srambical/test_extension/test/src/extension.ts",0,0,"",typescript,tab
+108,2372237,"/home/franz.srambical/test_extension/test/src/extension.ts",3150,0,"\n    // ---- Git extension state changes ----\n    try {\n        const gitExtension = vscode.extensions.getExtension('vscode.git')\n        if (gitExtension) {\n            if (!gitExtension.isActive) {\n                gitExtension.activate().then(() => setupGitListeners(gitExtension, log))\n            } else {\n                setupGitListeners(gitExtension, log)\n            }\n        }\n    } catch (e) {\n        log(`GIT_EXTENSION_ERROR: ${e}`)\n    }\n}\n\nfunction setupGitListeners(gitExtension: vscode.Extension<any>, log: (msg: string) => void) {\n    try {\n        const git = gitExtension.exports.getAPI(1)\n        if (!git) return\n\n        git.onDidChangeState(() => {\n            log(`GIT_STATE_CHANGE state=${git.state}`)\n        })\n\n        // Watch each repository\n        for (const repo of git.repositories) {\n            repo.state.onDidChange(() => {\n                log(\n                    `GIT_REPO_STATE_CHANGE\n  root=${repo.rootUri.toString()}\n  HEAD=${repo.state.HEAD?.name ?? 'detached'}\n  indexChanges=${repo.state.indexChanges.length}\n  workingTreeChanges=${repo.state.workingTreeChanges.length}`\n                )\n            })\n        }\n\n        // Watch for new repositories\n        git.onDidOpenRepository((repo: any) => {\n            log(`GIT_REPO_OPENED root=${repo.rootUri.toString()}`)\n            repo.state.onDidChange(() => {\n                log(\n                    `GIT_REPO_STATE_CHANGE\n  root=${repo.rootUri.toString()}\n  HEAD=${repo.state.HEAD?.name ?? 'detached'}\n  indexChanges=${repo.state.indexChanges.length}\n  workingTreeChanges=${repo.state.workingTreeChanges.length}`\n                )\n            })\n        })\n    } catch (e) {\n        log(`GIT_LISTENER_ERROR: ${e}`)\n    }\n",typescript,content
+109,2372238,"/home/franz.srambical/test_extension/test/src/extension.ts",2490,0,"    // ---- Document save ----\n    context.subscriptions.push(\n        vscode.workspace.onDidSaveTextDocument(doc => {\n            const uri = doc.uri\n            if (!isInWorkspace(uri)) return\n            if (shouldIgnore(uri)) return\n\n            log(\n                `DOC_SAVE\n  uri=${uri.toString()}\n  version=${doc.version}`\n            )\n        })\n    )\n\n    // ---- Selection changes ----\n    context.subscriptions.push(\n        vscode.window.onDidChangeTextEditorSelection(event => {\n            const uri = event.textEditor.document.uri\n            if (!isInWorkspace(uri)) return\n            if (shouldIgnore(uri)) return\n\n            const sel = event.selections[0]\n            const kindName = event.kind !== undefined\n                ? ['Keyboard', 'Mouse', 'Command'][event.kind] ?? 'Unknown'\n                : 'None'\n\n            log(\n                `SELECTION_CHANGE\n  uri=${uri.toString()}\n  kind=${kindName}\n  anchor=${sel.anchor.line}:${sel.anchor.character}\n  active=${sel.active.line}:${sel.active.character}\n  isEmpty=${sel.isEmpty}`\n            )\n        })\n    )\n\n    // ---- Active editor change (TAB) ----\n    context.subscriptions.push(\n        vscode.window.onDidChangeActiveTextEditor(editor => {\n            if (!editor) {\n                log(`ACTIVE_EDITOR_CHANGE editor=none`)\n                return\n            }\n            const uri = editor.document.uri\n            if (!isInWorkspace(uri)) return\n            if (shouldIgnore(uri)) return\n\n            log(\n                `ACTIVE_EDITOR_CHANGE\n  uri=${uri.toString()}\n  language=${editor.document.languageId}`\n            )\n        })\n    )\n\n    // ---- Terminal focus ----\n    context.subscriptions.push(\n        vscode.window.onDidChangeActiveTerminal(terminal => {\n            log(`TERMINAL_FOCUS name=${terminal?.name ?? 'none'}`)\n        })\n    )\n\n    // ---- Terminal shell execution (captures git commands run in terminal) ----\n    context.subscriptions.push(\n        vscode.window.onDidStartTerminalShellExecution(async event => {\n            const commandLine = event.execution.commandLine.value\n            log(\n                `TERMINAL_COMMAND_START\n  terminal=${event.terminal.name}\n  command=${commandLine}`\n            )\n\n            // Stream terminal output\n            const stream = event.execution.read()\n            for await (const data of stream) {\n                const preview = data.length > 200 ? data.slice(0, 200) + '...' : data\n                log(\n                    `TERMINAL_OUTPUT\n  terminal=${event.terminal.name}\n  data=${preview.replace(/\n/g, '\\n')}`\n                )\n            }\n        })\n    )\n\n",typescript,content
+110,2372238,"/home/franz.srambical/test_extension/test/src/extension.ts",1993,72,"  reason=${event.reason ?? 'unknown'}\n  active=${vscode.window.activeTextEditor?.document === event.document}\n${changeDetails}`",typescript,content
+111,2372238,"/home/franz.srambical/test_extension/test/src/extension.ts",1846,0,"            // Log detailed change info to help identify undo patterns\n            const changeDetails = event.contentChanges.map((c, i) => {\n                const textPreview = c.text.length > 50 ? c.text.slice(0, 50) + '...' : c.text\n                return `  change[${i}]: offset=${c.rangeOffset} len=${c.rangeLength} text=""${textPreview.replace(/\n/g, '\\n')}""`\n            }).join('\n')\n\n",typescript,content
+112,2372238,"/home/franz.srambical/test_extension/test/src/extension.ts",1585,38,"    // ---- Command execution tracking (undo/redo/git commands) ----\n    context.subscriptions.push(\n        vscode.commands.onDidExecuteCommand(event => {\n            // Log all commands if they're in our interesting set, or if they contain 'git' or 'undo' or 'redo'\n            const cmd = event.command\n            const isInteresting = INTERESTING_COMMANDS.has(cmd) ||\n                cmd.toLowerCase().includes('undo') ||\n                cmd.toLowerCase().includes('redo') ||\n                cmd.toLowerCase().includes('git')\n\n            if (isInteresting) {\n                log(\n                    `COMMAND_EXECUTED\n  command=${cmd}\n  args=${JSON.stringify(event.arguments ?? [])}`\n                )\n            }\n        })\n    )\n\n    // ---- Text document changes (with more detail for undo detection) ----",typescript,content
+113,2372238,"/home/franz.srambical/test_extension/test/src/extension.ts",677,0,"// Commands we care about for undo/redo/git tracking\nconst INTERESTING_COMMANDS = new Set([\n    'undo',\n    'redo',\n    'editor.action.undo',\n    'editor.action.redo',\n    'git.pull',\n    'git.pullRebase',\n    'git.push',\n    'git.stash',\n    'git.stashPop',\n    'git.stashApply',\n    'git.stashDrop',\n    'git.checkout',\n    'git.branch',\n    'git.merge',\n    'git.rebase',\n    'git.fetch',\n    'git.commit',\n    'git.commitStaged',\n    'git.stage',\n    'git.unstage',\n    'git.clean',\n    'git.reset',\n    'git.revert',\n])\n\n",typescript,content
+114,2381832,"/home/franz.srambical/test_extension/test/src/extension.ts",792,0,"",typescript,selection_mouse
+115,2381832,"/home/franz.srambical/test_extension/test/src/extension.ts",791,0,"",typescript,selection_command
+116,2389132,"/home/franz.srambical/test_extension/test/src/extension.ts",7534,1722,"",typescript,content
+117,2389133,"/home/franz.srambical/test_extension/test/src/extension.ts",4242,2632,"",typescript,content
+118,2389133,"/home/franz.srambical/test_extension/test/src/extension.ts",3690,127,"  active=${vscode.window.activeTextEditor?.document === event.document}`",typescript,content
+119,2389133,"/home/franz.srambical/test_extension/test/src/extension.ts",3150,393,"",typescript,content
+120,2389133,"/home/franz.srambical/test_extension/test/src/extension.ts",2111,816,"    // ---- Text document changes ----",typescript,content
+121,2389133,"/home/franz.srambical/test_extension/test/src/extension.ts",677,526,"",typescript,content
+122,2389179,"/home/franz.srambical/test_extension/test/src/extension.ts",3150,0,"\n    // ---- Git extension state changes ----\n    try {\n        const gitExtension = vscode.extensions.getExtension('vscode.git')\n        if (gitExtension) {\n            if (!gitExtension.isActive) {\n                gitExtension.activate().then(() => setupGitListeners(gitExtension, log))\n            } else {\n                setupGitListeners(gitExtension, log)\n            }\n        }\n    } catch (e) {\n        log(`GIT_EXTENSION_ERROR: ${e}`)\n    }\n}\n\nfunction setupGitListeners(gitExtension: vscode.Extension<any>, log: (msg: string) => void) {\n    try {\n        const git = gitExtension.exports.getAPI(1)\n        if (!git) return\n\n        git.onDidChangeState(() => {\n            log(`GIT_STATE_CHANGE state=${git.state}`)\n        })\n\n        // Watch each repository\n        for (const repo of git.repositories) {\n            repo.state.onDidChange(() => {\n                log(\n                    `GIT_REPO_STATE_CHANGE\n  root=${repo.rootUri.toString()}\n  HEAD=${repo.state.HEAD?.name ?? 'detached'}\n  indexChanges=${repo.state.indexChanges.length}\n  workingTreeChanges=${repo.state.workingTreeChanges.length}`\n                )\n            })\n        }\n\n        // Watch for new repositories\n        git.onDidOpenRepository((repo: any) => {\n            log(`GIT_REPO_OPENED root=${repo.rootUri.toString()}`)\n            repo.state.onDidChange(() => {\n                log(\n                    `GIT_REPO_STATE_CHANGE\n  root=${repo.rootUri.toString()}\n  HEAD=${repo.state.HEAD?.name ?? 'detached'}\n  indexChanges=${repo.state.indexChanges.length}\n  workingTreeChanges=${repo.state.workingTreeChanges.length}`\n                )\n            })\n        })\n    } catch (e) {\n        log(`GIT_LISTENER_ERROR: ${e}`)\n    }\n",typescript,content
+123,2389179,"/home/franz.srambical/test_extension/test/src/extension.ts",2490,0,"    // ---- Document save ----\n    context.subscriptions.push(\n        vscode.workspace.onDidSaveTextDocument(doc => {\n            const uri = doc.uri\n            if (!isInWorkspace(uri)) return\n            if (shouldIgnore(uri)) return\n\n            log(\n                `DOC_SAVE\n  uri=${uri.toString()}\n  version=${doc.version}`\n            )\n        })\n    )\n\n    // ---- Selection changes ----\n    context.subscriptions.push(\n        vscode.window.onDidChangeTextEditorSelection(event => {\n            const uri = event.textEditor.document.uri\n            if (!isInWorkspace(uri)) return\n            if (shouldIgnore(uri)) return\n\n            const sel = event.selections[0]\n            const kindName = event.kind !== undefined\n                ? ['Keyboard', 'Mouse', 'Command'][event.kind] ?? 'Unknown'\n                : 'None'\n\n            log(\n                `SELECTION_CHANGE\n  uri=${uri.toString()}\n  kind=${kindName}\n  anchor=${sel.anchor.line}:${sel.anchor.character}\n  active=${sel.active.line}:${sel.active.character}\n  isEmpty=${sel.isEmpty}`\n            )\n        })\n    )\n\n    // ---- Active editor change (TAB) ----\n    context.subscriptions.push(\n        vscode.window.onDidChangeActiveTextEditor(editor => {\n            if (!editor) {\n                log(`ACTIVE_EDITOR_CHANGE editor=none`)\n                return\n            }\n            const uri = editor.document.uri\n            if (!isInWorkspace(uri)) return\n            if (shouldIgnore(uri)) return\n\n            log(\n                `ACTIVE_EDITOR_CHANGE\n  uri=${uri.toString()}\n  language=${editor.document.languageId}`\n            )\n        })\n    )\n\n    // ---- Terminal focus ----\n    context.subscriptions.push(\n        vscode.window.onDidChangeActiveTerminal(terminal => {\n            log(`TERMINAL_FOCUS name=${terminal?.name ?? 'none'}`)\n        })\n    )\n\n    // ---- Terminal shell execution (captures git commands run in terminal) ----\n    context.subscriptions.push(\n        vscode.window.onDidStartTerminalShellExecution(async event => {\n            const commandLine = event.execution.commandLine.value\n            log(\n                `TERMINAL_COMMAND_START\n  terminal=${event.terminal.name}\n  command=${commandLine}`\n            )\n\n            // Stream terminal output\n            const stream = event.execution.read()\n            for await (const data of stream) {\n                const preview = data.length > 200 ? data.slice(0, 200) + '...' : data\n                log(\n                    `TERMINAL_OUTPUT\n  terminal=${event.terminal.name}\n  data=${preview.replace(/\n/g, '\\n')}`\n                )\n            }\n        })\n    )\n\n",typescript,content
+124,2389179,"/home/franz.srambical/test_extension/test/src/extension.ts",1993,72,"  reason=${event.reason ?? 'unknown'}\n  active=${vscode.window.activeTextEditor?.document === event.document}\n${changeDetails}`",typescript,content
+125,2389179,"/home/franz.srambical/test_extension/test/src/extension.ts",1846,0,"            // Log detailed change info to help identify undo patterns\n            const changeDetails = event.contentChanges.map((c, i) => {\n                const textPreview = c.text.length > 50 ? c.text.slice(0, 50) + '...' : c.text\n                return `  change[${i}]: offset=${c.rangeOffset} len=${c.rangeLength} text=""${textPreview.replace(/\n/g, '\\n')}""`\n            }).join('\n')\n\n",typescript,content
+126,2389179,"/home/franz.srambical/test_extension/test/src/extension.ts",1585,38,"    // NOTE: vscode.commands.onDidExecuteCommand is NOT available in the stable API.\n    // Undo/redo operations will appear as TEXT_CHANGE events with reason=1 (Undo) or reason=2 (Redo).\n    // Git commands via VS Code UI will trigger GIT_REPO_STATE_CHANGE events.\n    // Git commands via terminal will trigger TERMINAL_COMMAND_START events.\n\n    // ---- Text document changes (with more detail for undo detection) ----",typescript,content
+127,2389179,"/home/franz.srambical/test_extension/test/src/extension.ts",677,0,"// Commands we care about for undo/redo/git tracking\nconst INTERESTING_COMMANDS = new Set([\n    'undo',\n    'redo',\n    'editor.action.undo',\n    'editor.action.redo',\n    'git.pull',\n    'git.pullRebase',\n    'git.push',\n    'git.stash',\n    'git.stashPop',\n    'git.stashApply',\n    'git.stashDrop',\n    'git.checkout',\n    'git.branch',\n    'git.merge',\n    'git.rebase',\n    'git.fetch',\n    'git.commit',\n    'git.commitStaged',\n    'git.stage',\n    'git.unstage',\n    'git.clean',\n    'git.reset',\n    'git.revert',\n])\n\n",typescript,content
+128,2441118,"/home/franz.srambical/test_extension/test/src/extension.ts",0,8862,"import * as vscode from 'vscode'\n\nlet output: vscode.OutputChannel\nlet tracingEnabled = false\n\nfunction isInWorkspace(uri: vscode.Uri): boolean {\n    const folders = vscode.workspace.workspaceFolders\n    if (!folders) return false\n    return folders.some(f => uri.fsPath.startsWith(f.uri.fsPath))\n}\n\nfunction shouldIgnore(uri: vscode.Uri): boolean {\n    const p = uri.fsPath\n\n    // Ignore VS Code internals and extension noise\n    if (p.includes(`${pathSep}.vscode${pathSep}`)) return true\n    if (p.includes(`${pathSep}node_modules${pathSep}`)) return true\n    if (p.includes(`${pathSep}.git${pathSep}`)) return true\n\n    return false\n}\n\nconst pathSep = require('path').sep\n\nexport function activate(context: vscode.ExtensionContext) {\n    output = vscode.window.createOutputChannel('Edit Trace Probe')\n\n    const log = (msg: string) => {\n        if (!tracingEnabled) return\n        output.appendLine(`[${new Date().toISOString()}] ${msg}`)\n    }\n\n    // ---- Commands ----\n    context.subscriptions.push(\n        vscode.commands.registerCommand('editTraceProbe.start', () => {\n            tracingEnabled = true\n            output.show(true)\n            output.appendLine('--- TRACE STARTED ---')\n        })\n    )\n\n    context.subscriptions.push(\n        vscode.commands.registerCommand('editTraceProbe.stop', () => {\n            output.appendLine('--- TRACE STOPPED ---')\n            tracingEnabled = false\n        })\n    )\n\n    context.subscriptions.push(\n        vscode.commands.registerCommand('editTraceProbe.showOutput', () => {\n            output.show(true)\n        })\n    )\n\n    // ---- Text document changes ----\n    context.subscriptions.push(\n        vscode.workspace.onDidChangeTextDocument(event => {\n            const uri = event.document.uri\n            if (!isInWorkspace(uri)) return\n            if (shouldIgnore(uri)) return\n\n            log(\n                `TEXT_CHANGE\n  uri=${uri.toString()}\n  version=${event.document.version}\n  changes=${event.contentChanges.length}\n  active=${vscode.window.activeTextEditor?.document === event.document}`\n            )\n        })\n    )\n\n    // ---- Document open ----\n    context.subscriptions.push(\n        vscode.workspace.onDidOpenTextDocument(doc => {\n            const uri = doc.uri\n            if (!isInWorkspace(uri)) return\n            if (shouldIgnore(uri)) return\n\n            log(\n                `DOC_OPEN\n  uri=${uri.toString()}\n  version=${doc.version}\n  language=${doc.languageId}`\n            )\n        })\n    )\n\n    // ---- Filesystem watcher (workspace only) ----\n    const watcher = vscode.workspace.createFileSystemWatcher('**/*')\n    context.subscriptions.push(watcher)\n\n    watcher.onDidChange(uri => {\n        if (!isInWorkspace(uri)) return\n        if (shouldIgnore(uri)) return\n        log(`FS_CHANGE uri=${uri.toString()}`)\n    })\n\n    watcher.onDidCreate(uri => {\n        if (!isInWorkspace(uri)) return\n        if (shouldIgnore(uri)) return\n        log(`FS_CREATE uri=${uri.toString()}`)\n    })\n\n    watcher.onDidDelete(uri => {\n        if (!isInWorkspace(uri)) return\n        if (shouldIgnore(uri)) return\n        log(`FS_DELETE uri=${uri.toString()}`)\n    })\n}\n",typescript,content
+129,2457935,"/home/franz.srambical/test_extension/test/src/extension.ts",374,0,"",typescript,selection_mouse
+130,2457942,"/home/franz.srambical/test_extension/test/src/extension.ts",373,0,"",typescript,selection_command
+131,2486524,"/home/franz.srambical/test_extension/test/src/extension.ts",375,0,"",typescript,selection_command
+132,2486622,"/home/franz.srambical/test_extension/test/src/extension.ts",399,0,"",typescript,selection_command
+133,2486789,"/home/franz.srambical/test_extension/test/src/extension.ts",451,0,"",typescript,selection_command
+134,2486905,"/home/franz.srambical/test_extension/test/src/extension.ts",514,0,"",typescript,selection_command
+135,2487568,"/home/franz.srambical/test_extension/test/src/extension.ts",3152,0,"",typescript,selection_command
+136,2488265,"/home/franz.srambical/test_extension/test/src/extension.ts",3150,0,"",typescript,selection_command
+137,2488520,"/home/franz.srambical/test_extension/test/src/extension.ts",3143,0,"",typescript,selection_command
+138,2488548,"/home/franz.srambical/test_extension/test/src/extension.ts",3096,0,"",typescript,selection_command
+139,2488583,"/home/franz.srambical/test_extension/test/src/extension.ts",3058,0,"",typescript,selection_command
+140,2488606,"/home/franz.srambical/test_extension/test/src/extension.ts",3018,0,"",typescript,selection_command
+141,2488647,"/home/franz.srambical/test_extension/test/src/extension.ts",2985,0,"",typescript,selection_command
+142,2488682,"/home/franz.srambical/test_extension/test/src/extension.ts",2984,0,"",typescript,selection_command
+143,2488714,"/home/franz.srambical/test_extension/test/src/extension.ts",2977,0,"",typescript,selection_command
+144,2488748,"/home/franz.srambical/test_extension/test/src/extension.ts",2930,0,"",typescript,selection_command
+145,2488781,"/home/franz.srambical/test_extension/test/src/extension.ts",2892,0,"",typescript,selection_command
+146,2488805,"/home/franz.srambical/test_extension/test/src/extension.ts",2852,0,"",typescript,selection_command
+147,2488849,"/home/franz.srambical/test_extension/test/src/extension.ts",2819,0,"",typescript,selection_command
+148,2488881,"/home/franz.srambical/test_extension/test/src/extension.ts",2818,0,"",typescript,selection_command
+149,2488915,"/home/franz.srambical/test_extension/test/src/extension.ts",2811,0,"",typescript,selection_command
+150,2488947,"/home/franz.srambical/test_extension/test/src/extension.ts",2764,0,"",typescript,selection_command
+151,2488980,"/home/franz.srambical/test_extension/test/src/extension.ts",2726,0,"",typescript,selection_command
+152,2489005,"/home/franz.srambical/test_extension/test/src/extension.ts",2686,0,"",typescript,selection_command
+153,2489046,"/home/franz.srambical/test_extension/test/src/extension.ts",2653,0,"",typescript,selection_command
+154,2489120,"/home/franz.srambical/test_extension/test/src/extension.ts",2652,0,"",typescript,selection_command
+155,2489120,"/home/franz.srambical/test_extension/test/src/extension.ts",2612,0,"",typescript,selection_command
+156,2489173,"/home/franz.srambical/test_extension/test/src/extension.ts",2543,0,"",typescript,selection_command
+157,2489178,"/home/franz.srambical/test_extension/test/src/extension.ts",2490,0,"",typescript,selection_command
+158,2489222,"/home/franz.srambical/test_extension/test/src/extension.ts",2489,0,"",typescript,selection_command
+159,2489248,"/home/franz.srambical/test_extension/test/src/extension.ts",2483,0,"",typescript,selection_command
+160,2489282,"/home/franz.srambical/test_extension/test/src/extension.ts",2472,0,"",typescript,selection_command
+161,2495724,"/home/franz.srambical/test_extension/test/src/extension.ts",1080,0,"",typescript,selection_keyboard
+162,2497038,"/home/franz.srambical/test_extension/test/src/extension.ts",0,0,"",typescript,selection_keyboard
+163,2497602,"/home/franz.srambical/test_extension/test/src/extension.ts",1460,0,"",typescript,selection_keyboard
+164,2497786,"/home/franz.srambical/test_extension/test/src/extension.ts",2852,0,"",typescript,selection_keyboard
+165,2497911,"/home/franz.srambical/test_extension/test/src/extension.ts",3152,0,"",typescript,selection_keyboard
+166,2506459,"/home/franz.srambical/test_extension/test/src/extension.ts",2659,0,"",typescript,selection_mouse
+167,2529907,"/home/franz.srambical/test_extension/test/src/extension.ts",2665,0,"",typescript,selection_command
+168,2533738,"/home/franz.srambical/test_extension/test/src/extension.ts",1681,0,"",typescript,selection_command
+169,2557502,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",0,0,"",Log,tab
+170,2582392,"/home/franz.srambical/test_extension/test/src/extension.ts",0,0,"",typescript,tab
+171,2584596,"crates/cli/src/main.rs",0,0,"",rust,tab
+172,2589664,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",0,0,"",Log,tab
+173,2589664,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",824,0,"",Log,selection_mouse
+174,2590108,"crates/cli/src/main.rs",0,0,"",rust,tab
+175,2592126,"crates/cli/src/main.rs",10299,0,"\n",rust,content
+176,2593178,"crates/cli/src/main.rs",10300,0,"t",rust,content
+177,2593179,"crates/cli/src/main.rs",10301,0,"",rust,selection_keyboard
+178,2593271,"crates/cli/src/main.rs",10301,0,"e",rust,content
+179,2593272,"crates/cli/src/main.rs",10302,0,"",rust,selection_keyboard
+180,2593461,"crates/cli/src/main.rs",10302,0,"s",rust,content
+181,2593461,"crates/cli/src/main.rs",10303,0,"",rust,selection_keyboard
+182,2593472,"crates/cli/src/main.rs",10303,0,"t",rust,content
+183,2593472,"crates/cli/src/main.rs",10304,0,"",rust,selection_keyboard
+184,2594710,"crates/cli/src/main.rs",10303,0,"",rust,selection_command
+185,2600450,"crates/cli/src/main.rs",10299,5,"",rust,content
+186,2604337,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",0,0,"",Log,tab
+187,2608424,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",1131,0,"",Log,selection_mouse
+188,2608584,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",1131,1,"1",Log,selection_mouse
+189,2608742,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",1121,12,"  changes=1\n",Log,selection_mouse
+190,2609817,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",1007,0,"",Log,selection_mouse
+191,2609984,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",1006,2,"44",Log,selection_mouse
+192,2610148,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",988,39,"[2026-01-07T11:30:44.740Z] TEXT_CHANGE\n",Log,selection_mouse
+193,2610511,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",988,120,"[2026-01-07T11:30:44.740Z] TEXT_CHANGE\n  uri=file:///home/franz.srambical/crowd-pilot-serializer/crates/cli/src/main.rs\n",Log,selection_mouse
+194,2610533,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",988,133,"[2026-01-07T11:30:44.740Z] TEXT_CHANGE\n  uri=file:///home/franz.srambical/crowd-pilot-serializer/crates/cli/src/main.rs\n  version=12\n",Log,selection_mouse
+195,2610583,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",988,145,"[2026-01-07T11:30:44.740Z] TEXT_CHANGE\n  uri=file:///home/franz.srambical/crowd-pilot-serializer/crates/cli/src/main.rs\n  version=12\n  changes=1\n",Log,selection_mouse
+196,2610618,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",988,159,"[2026-01-07T11:30:44.740Z] TEXT_CHANGE\n  uri=file:///home/franz.srambical/crowd-pilot-serializer/crates/cli/src/main.rs\n  version=12\n  changes=1\n  active=true\n",Log,selection_mouse
+197,2610666,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",988,198,"[2026-01-07T11:30:44.740Z] TEXT_CHANGE\n  uri=file:///home/franz.srambical/crowd-pilot-serializer/crates/cli/src/main.rs\n  version=12\n  changes=1\n  active=true\n[2026-01-07T11:30:44.741Z] TEXT_CHANGE\n",Log,selection_mouse
+198,2610720,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",988,279,"[2026-01-07T11:30:44.740Z] TEXT_CHANGE\n  uri=file:///home/franz.srambical/crowd-pilot-serializer/crates/cli/src/main.rs\n  version=12\n  changes=1\n  active=true\n[2026-01-07T11:30:44.741Z] TEXT_CHANGE\n  uri=file:///home/franz.srambical/crowd-pilot-serializer/crates/cli/src/main.rs\n",Log,selection_mouse
+199,2610765,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",988,292,"[2026-01-07T11:30:44.740Z] TEXT_CHANGE\n  uri=file:///home/franz.srambical/crowd-pilot-serializer/crates/cli/src/main.rs\n  version=12\n  changes=1\n  active=true\n[2026-01-07T11:30:44.741Z] TEXT_CHANGE\n  uri=file:///home/franz.srambical/crowd-pilot-serializer/crates/cli/src/main.rs\n  version=12\n",Log,selection_mouse
+200,2610883,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",988,304,"[2026-01-07T11:30:44.740Z] TEXT_CHANGE\n  uri=file:///home/franz.srambical/crowd-pilot-serializer/crates/cli/src/main.rs\n  version=12\n  changes=1\n  active=true\n[2026-01-07T11:30:44.741Z] TEXT_CHANGE\n  uri=file:///home/franz.srambical/crowd-pilot-serializer/crates/cli/src/main.rs\n  version=12\n  changes=0\n",Log,selection_mouse
+201,2611801,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",988,318,"[2026-01-07T11:30:44.740Z] TEXT_CHANGE\n  uri=file:///home/franz.srambical/crowd-pilot-serializer/crates/cli/src/main.rs\n  version=12\n  changes=1\n  active=true\n[2026-01-07T11:30:44.741Z] TEXT_CHANGE\n  uri=file:///home/franz.srambical/crowd-pilot-serializer/crates/cli/src/main.rs\n  version=12\n  changes=0\n  active=true\n",Log,selection_mouse
+202,2612892,"extension-output-undefined_publisher.test-#1-Edit Trace Probe",1279,0,"",Log,selection_mouse
+203,2618417,"crates/cli/src/main.rs",0,0,"",rust,tab